Tianhao Wu, Weizhe Yuan, Olga Golovneva, Jing Xu, Yuandong Tian, Jiantao Jiao, Jason E Weston, Sainbayar Sukhbaatar. Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing. 2025.

Large Language Models (LLMs), while demonstrating remarkable capabilities across various applications, present significant challenges during inference due to their substantial model size, especially when deployed on edge devices. Activation sparsity offers a promising solution to reduce computation and memory movement, enabling more efficient inference, particularly for small-batch on-device applications. However, current approaches face limitations with non-ReLU activation function, which are foundational to most advanced LLMs, or require heavy continual training. Additionally, the difficulty in predicting active channels and limited achievable sparsity ratios constrain the effectiveness of activation sparsity-based methods. In this paper, we introduce R-Sparse, a training-free activation sparsity approach capable of achieving high sparsity levels in advanced LLMs. We conducted two preliminary investigations into how different components contribute to the output within a single linear layer and found two key observations: (i) the non-sparse components of the input function can be regarded as a few bias terms, and (ii) The full computation can be effectively approximated by an appropriate combination of input channels and weight singular values. Building on this, we replace the linear layers in LLMs with a rank-aware sparse inference method that leverages the sparsity of input channels and singular value components, eliminating the need for active channel prediction like the output sparsity based approaches. Experiments on Llama-2/3 and Mistral models across ten diverse tasks demonstrate that R-Sparse achieves comparable performance at 50\% model-level sparsity, resulting in a significant 43\% end-to-end efficient improvements with customized kernels.

Large language models (LLMs) with long context windows have gained significant attention. However, the KV cache, stored to avoid re-computation, becomes a bottleneck. Various dynamic sparse or TopK-based attention approximation methods have been proposed to leverage the common insight that attention is sparse. In this paper, we first show that TopK attention itself suffers from quality degradation in certain downstream tasks because attention is not always as sparse as expected. Rather than selecting the keys and values with the highest attention scores, sampling with theoretical guarantees can provide a better estimation for attention output. To make the sampling-based approximation practical in LLM generation, we propose MagicPIG, a heterogeneous system based on Locality Sensitive Hashing (LSH). MagicPIG significantly reduces the workload of attention computation while preserving high accuracy for diverse tasks. MagicPIG stores the LSH hash tables and runs the attention computation on the CPU, which allows it to serve longer contexts and larger batch sizes with high approximation accuracy. MagicPIG can improve decoding throughput by up to $5\times$ across various GPU hardware and achieve 54ms decoding latency on a single RTX 4090 for Llama-3.1-8B-Instruct model with a context of 96k tokens.

Reinforcement learning (RL) promises a framework for near-universal problem-solving. In practice however, RL algorithms are often tailored to specific benchmarks, relying on carefully tuned hyperparameters and algorithmic choices. Recently, powerful model-based RL methods have shown impressive general results across benchmarks but come at the cost of increased complexity and slow run times, limiting their broader applicability. In this paper, we attempt to find a unifying model-free deep RL algorithm that can address a diverse class of domains and problem settings. To achieve this, we leverage model-based representations that approximately linearize the value function, taking advantage of the denser task objectives used by model-based RL while avoiding the costs associated with planning or simulated trajectories. We evaluate our algorithm, MR.Q, on a variety of common RL benchmarks with a single set of hyperparameters and show a competitive performance against domain-specific and general baselines, providing a concrete step towards building general-purpose model-free deep RL algorithms.

Post-training quantization (PTQ) techniques applied to weights, activations, and the KV cache greatly reduce memory usage, latency, and power consumption of Large Language Models (LLMs), but may lead to large quantization errors when outliers are present. Rotating activation or weight matrices helps remove outliers and benefits quantization. In this work, we identify a collection of applicable rotation parameterizations that lead to identical outputs in full-precision Transformer architectures while enhancing quantization accuracy. In addition, we find that some random rotations lead to much better quantization than others, with an up to 13 points difference in downstream zero-shot reasoning performance. As a result, we propose SpinQuant, a novel approach that incorporates learned rotation matrices for optimal quantized network accuracy. With 4-bit quantization of weight, activation, and KV-cache, SpinQuant narrows the accuracy gap on zero-shot reasoning tasks with full precision to merely 2.9 points on the LLaMA-2 7B model, surpassing LLM-QAT by 19.1 points and SmoothQuant by 25.0 points. Furthermore, SpinQuant also outperforms concurrent work QuaRot, which applies random rotations to remove outliers. In particular, for LLaMA-3 8B models that are hard to quantize, SpinQuant reduces the gap to full precision by up to 45.1% relative to QuaRot. Code is available at https://github.com/facebookresearch/SpinQuant.

Aligning AI systems with human preferences typically suffers from the infamous *reward hacking* problem, where optimization of an imperfect reward model leads to undesired behaviors. In this paper, we investigate reward hacking in offline preference optimization, which aims to improve an initial model using a preference dataset. We identify two types of reward hacking stemming from statistical fluctuations in the dataset: Type I Reward Hacking due to subpar choices appearing more favorable, and Type II Reward Hacking due to decent choices appearing less desirable. We prove that many (mainstream or theoretical) preference optimization methods suffer from both types of reward hacking. To mitigate Type I Reward Hacking, we propose POWER, a new preference optimization method that combines Guiasu's weighted entropy with a robust reward maximization objective. POWER enjoys finite-sample guarantees under general function approximation, competing with the best covered policy in the data. To mitigate Type II Reward Hacking, we analyze the learning dynamics of preference optimization and develop a novel technique that dynamically updates preference labels toward certain "stationary labels", resulting in diminishing gradients for untrustworthy samples. Empirically, POWER with dynamic labels (DL) consistently outperforms state-of-the-art methods on alignment benchmarks, achieving improvements of up to **13.0** points on AlpacaEval 2 and **11.5** points on Arena-Hard over DPO, while also improving or maintaining performance on downstream tasks such as mathematical reasoning. Strong theoretical guarantees and empirical results demonstrate the promise of POWER-DL in mitigating reward hacking.

In cognition theory, human thinking is governed by two systems: the fast and intuitive System 1 and the slower but more deliberative System 2. Analogously, Large Language Models (LLMs) can operate in two reasoning modes: outputting only the solutions (\emph{fast mode}) or both the reasoning chain and the final solution (\emph{slow mode}). We present \dualformer, a single Transformer model that seamlessly integrates both the fast and slow reasoning modes by training on randomized reasoning traces, where different parts of the traces are strategically dropped during training. At inference time, \dualformer can be easily configured to execute in either fast or slow mode, or automatically decide which mode to engage (\emph{auto mode}). It outperforms baselines in both performance and computational efficiency across all three modes: \textbf{(1)} in slow mode, \dualformer achieves $97.6\%$ optimal rate on unseen $30 \times 30$ maze tasks, surpassing the \searchformer baseline (93.3\%) trained on data with complete reasoning traces, with $45.5\%$ fewer reasoning steps; \textbf{(2)} in fast mode, \dualformer achieves $80\%$ optimal rate, significantly outperforming the Solution-Only model trained on solution-only data, which has an optimal rate of only 30\%; \textbf{(3)} in auto mode, \dualformer achieves $96.6\%$ optimal rate with $59.9\%$ fewer steps than \searchformer. For math reasoning problems, our techniques have also achieved improved performance with LLM fine-tuning, demonstrating its generalization beyond task-specific models. We open source our code at https://github.com/facebookresearch/dualformer.

Large Language Models (LLMs) matrices can often be expressed in low-rank format with potential to relax memory and compute resource requirements. Unlike previous works which pivot around developing novel matrix decomposition algorithms, in this work we focus to study the emerging non-uniform low-rank properties across weight matrices in LLMs through the lens of stabilizing gradient subspace. \textit{Firstly,} we provide a theoretical framework to understand the stabilization of gradient subspaces through Hessian analysis. \textit{Secondly,} we empirically establish a consequential relationship between the gradient dynamics and low-rank expressiveness of weight matrices. Our findings reveal that different LLM components exhibit varying levels of converged low-rank structure, necessitating a non-uniform rank reduction across them to minimize performance drop due to compression. In view of that, we present \textit{Weight Low-Rank Projection} \textbf{(WeLore)} that unifies weight compression and memory-efficient fine-tuning as ONE, in a data-agnostic and one-shot way. Going beyond only as a compression technique, WeLore categorizes weight matrices into Low-rank Components (LRCs) and Non-Low-rank Components (N-LRCs) based on their ability to express themselves as low-rank. Our gradient dynamics perspective illustrate that \textit{LRCs tend to have better finetuning capabilities} and their standalone finetuning can closely mimic (sometimes outperform) the training loss trajectory and performance of full-finetuning with notable memory and compute footprint reduction. All codes and checkpoints will be released.

Large Language Models (LLMs) are vulnerable to **jailbreaking attacks** that lead to generation of inappropriate or harmful content. Manual red-teaming requires a time-consuming search for adversarial prompts, whereas automatic adversarial prompt generation often leads to semantically meaningless attacks that do not scale well. In this paper, we present a novel method that uses another LLM, called **AdvPrompter**, to generate human-readable adversarial prompts in seconds. AdvPrompter, which is trained using an alternating optimization algorithm, generates suffixes that veil the input instruction without changing its meaning, such that the TargetLLM is lured to give a harmful response. Experimental results on popular open source TargetLLM show highly competitive results on the AdvBench and HarmBench datasets, that also transfer to closed-source black-box LLMs. We also show that training on adversarial suffixes generated by AdvPrompter is a promising strategy for improving the robustness of LLMs to jailbreaking attacks.

Large Language Models (LLMs) excel at reasoning and planning when trained on chain-of-thought (CoT) data, where the step-by-step thought process is explicitly outlined by text tokens. However, this results in lengthy inputs where many words support textual coherence rather than core reasoning information, and processing these inputs consumes substantial computation resources. In this work, we propose a hybrid representation of the reasoning process, where we partially abstract away the initial reasoning steps using latent discrete tokens generated by VQ-VAE, significantly reducing the length of reasoning traces. We explore the use of latent trace abstractions in two scenarios: 1) training the model from scratch for the Keys-Finding Maze problem, 2) fine-tuning LLMs on this hybrid data with an extended vocabulary including unseen latent tokens, for both logical and mathematical reasoning problems. To facilitate effective learning, we introduce a simple training procedure that randomly mixes latent and text tokens, which enables fast adaptation to new latent tokens. Our approach consistently outperforms the baselines methods in various benchmarks, such as Math (+4.2%, Llama-3.2-1B), GSM8K (+4.1%, Llama-3.2-3B), and Fresh-Gaokao-Math-2023 (+13.3%, Llama-3.1-8B) with an average reduction of 17% in reasoning trace’s length.

Recently, long-context large language models (LLMs) have shown strong performance in information retrieval and long-document QA. However, to tackle the most challenging intellectual problems, LLMs must reason effectively in long and complex contexts (e.g., frontier mathematical research). Studying how LLMs handle increasing reasoning complexity and context length is essential, yet existing benchmarks lack a solid basis for quantitative evaluation. Inspired by the abstraction of GSM-8K problems as computational graphs—and the ability to introduce noise by adding unnecessary nodes and edges—we develop a grade-school math problem generator capable of producing arithmetic problems with infinite difficulty and context length under fine-grained control. Using our newly synthesized GSM-$\infty$ benchmark, we comprehensively evaluate existing LLMs. We find a consistent sigmoid decline in reasoning performance as complexity increases, along with a systematic inference scaling trend: exponentially increasing inference computation yields only linear performance gains. These findings underscore the fundamental limitations of current long-context LLMs and the key challenges in scaling reasoning capabilities. Our GSM-$\infty$ benchmark provides a scalable and controllable testbed for systematically studying and advancing LLM reasoning in long and complex contexts.

Contemporary evaluation techniques are inadequate for agentic systems. These approaches either focus exclusively on final outcomes—ignoring the step-by-step nature of the thinking done by agentic systems—or require excessive manual labour. To address this, we introduce the Agent-as-a-Judge framework, wherein agentic systems are used to evaluate agentic systems. This is a natural extension of the LLM-as-a-Judge framework, incorporating agentic features that enable intermediate feedback for the entire task-solving processes for more precise evaluations. We apply the Agent-as-a-Judge framework to the task of code generation. To overcome issues with existing benchmarks and provide a proof-of-concept testbed for Agent-as-a-Judge, we present DevAI, a new benchmark of 55 realistic AI code generation tasks. DevAI includes rich manual annotations, like a total of 365 hierarchical solution requirements, which make it particularly suitable for an agentic evaluator. We benchmark three of the top code-generating agentic systems using Agent-as-a-Judge and find that our framework dramatically outperforms LLM-as-a-Judge and is as reliable as our human evaluation baseline. Altogether, we believe that this work represents a concrete step towards enabling vastly more sophisticated agentic systems. To help that, our dataset and the full implementation of Agent-as-a-Judge will be publically available at https://github.com/metauto-ai/agent-as-a-judge

The optimal bit-width for achieving the best trade-off between quantized model size and accuracy has been a subject of ongoing debate. While some advocate for 4-bit quantization, others propose that 1.58-bit offers superior results. However, the lack of a cohesive framework for different bits has left such conclusions relatively tenuous. We present ParetoQ, the first unified framework that facilitates rigorous comparisons across 1-bit, 1.58-bit, 2-bit, 3-bit, and 4-bit quantization settings. Our findings reveal a notable learning transition between 2 and 3 bits: For 3-bits and above, the fine-tuned models stay close to their original pre-trained distributions, whereas for learning 2-bit networks or below, the representations change drastically. By optimizing training schemes and refining quantization functions, ParetoQ surpasses all previous methods tailored to specific bit widths. Remarkably, our ParetoQ ternary 600M-parameter model even outperforms the previous SoTA ternary 3B-parameter model in accuracy, using only one-fifth of the parameters. Extensive experimentation shows that ternary, 2-bit, and 3-bit quantization maintains comparable performance in the size-accuracy trade-off and generally exceeds 4-bit and binary quantization. Considering hardware constraints, 2-bit quantization offers promising potential for memory reduction and speedup.

We prove rich algebraic structures of the solution space for 2-layer neural networks with quadratic activation and $L_2$ loss, trained on reasoning tasks in Abelian group (e.g., modular addition). Such a rich structure enables \emph{analytical} construction of global optimal solutions from partial solutions that only satisfy part of the loss, despite its high nonlinearity. We coin the framework as \ours{} (\emph{\underline{Co}mposing \underline{G}lobal \underline{S}olutions}). Specifically, we show that the weight space over different numbers of hidden nodes of the 2-layer network is equipped with a semi-ring algebraic structure, and the loss function to be optimized consists of \emph{sum potentials}, which are ring homomorphisms, allowing partial solutions to be composed into global ones by ring addition and multiplication. Our experiments show that around $95\%$ of the solutions obtained by gradient descent match exactly our theoretical constructions. Although the global solutions constructed only required a small number of hidden nodes, our analysis on gradient dynamics shows that overparameterization asymptotically decouples training dynamics and is beneficial. We further show that training dynamics favors simpler solutions under weight decay, and thus high-order global solutions such as perfect memorization are unfavorable. The code is open sourced\footnote{\url{https://github.com/facebookresearch/luckmatters/tree/yuandong3/ssl/real-dataset}}.

Scaling reasoning capabilities beyond traditional domains such as math and coding is hindered by the lack of diverse and high-quality questions. To overcome this limitation, we introduce a scalable approach for generating diverse and challenging reasoning questions, accompanied by reference answers. We present NaturalReasoning, a comprehensive dataset comprising 2.8 million questions that span multiple domains, including STEM fields (e.g., Physics, Computer Science), Economics, Social Sciences, and more. We demonstrate the utility of the questions in NaturalReasoning through knowledge distillation experiments which show that NaturalReasoning can effectively elicit and transfer reasoning capabilities from a strong teacher model. Furthermore, we demonstrate that NaturalReasoning is also effective for unsupervised self-training using external reward models or self-rewarding.

Many jailbreak attacks on large language models (LLMs) rely on a common objective: making the model respond with the prefix ``Sure, here is (harmful request)''. While straightforward, this objective has two limitations: limited control over model behaviors, yielding incomplete or unrealistic jailbroken responses, and a rigid format that hinders optimization. We introduce AdvPrefix, a plug-and-play prefix-forcing objective that selects one or more model-dependent prefixes by combining two criteria: high prefilling attack success rates and low negative log-likelihood. AdvPrefix integrates seamlessly into existing jailbreak attacks to mitigate the previous limitations for free. For example, replacing GCG's default prefixes on Llama-3 improves nuanced attack success rates from 14\% to 80\%, revealing that current safety alignment fails to generalize to new prefixes. Code and selected prefixes are released.

Large Language Models (LLMs) have demonstrated remarkable performance in many applications, including challenging reasoning problems via chain-of-thought (CoT) techniques that generate ``thinking tokens'' before answering the questions. While existing theoretical works demonstrate that CoT with discrete tokens boosts the capability of LLMs, recent work on continuous CoT lacks a theoretical understanding of why it outperforms discrete counterparts in various reasoning tasks, such as directed graph reachability, a fundamental graph reasoning problem that includes many practical domain applications as special cases. In this paper, we prove that a two-layer transformer with $D$ steps of continuous CoT can solve the directed graph reachability problem, where $D$ is the diameter of the graph, while the best known result of constant-depth transformers with discrete CoT requires $O(n^2)$ decoding steps where $n$ is the number of vertices ($D<n$). In our construction, each continuous thought vector is a superposition state that encodes multiple search frontiers simultaneously (i.e., parallel breadth-first search (BFS)), while discrete CoT must choose a single path sampled from the superposition state, which leads to a sequential search that requires many more steps and may be trapped in local solutions. We also performed extensive experiments to verify that our theoretical construction aligns well with the empirical solution obtained via training dynamics. Notably, encoding of multiple search frontiers as a superposition state automatically emerges in training continuous CoT, without explicit supervision to guide the model to explore multiple paths simultaneously.

Recent research has increasingly focused on evaluating large language models' (LLMs) alignment with diverse human values and preferences, particularly for open-ended tasks like story generation.Traditional evaluation metrics rely heavily on lexical similarity with humanwritten references, often showing poor correlation with human judgments and failing to account for alignment with the diversity of human preferences.To address these challenges, we introduce PERSE, an interpretable evaluation framework designed to assess alignment with specific human preferences.It is tuned to infer specific preferences from an in-context personal profile and evaluate the alignment between the generated content and personal preferences.PERSE enhances interpretability by providing detailed comments and fine-grained scoring, facilitating more personalized content generation.Our 13B LLaMA-2-based PERSE shows a 15.8% increase in Kendall correlation and a 13.7% rise in accuracy with zero-shot reviewers compared to GPT-4.It also outperforms GPT-4 by 46.01% in Kendall correlation on new domains, indicating its transferability 1 .

Humanoid control is an important research challenge offering avenues for integration into human-centric infrastructures and enabling physics-driven humanoid animations. The daunting challenges in this field stem from the difficulty of optimizing in high-dimensional action spaces and the instability introduced by the bipedal morphology of humanoids. However, the extensive collection of human motion-captured data and the derived datasets of humanoid trajectories, such as MoCapAct, paves the way to tackle these challenges. In this context, we present Humanoid Generalist Autoencoding Planner (H-GAP), a state-action trajectory generative model trained on humanoid trajectories derived from human motion-captured data, capable of adeptly handling downstream control tasks with Model Predictive Control (MPC). For 56 degrees of freedom humanoid, we empirically demonstrate that H-GAP learns to represent and generate a wide range of motor behaviors. Further, without any learning from online interactions, it can also flexibly transfer these behaviours to solve novel downstream control tasks via planning. Notably, H-GAP excels established MPC baselines with access to the ground truth model, and is superior or comparable to offline RL methods trained for individual tasks. Finally, we do a series of empirical studies on the scaling properties of H-GAP, showing the potential for performance gains via additional data but not computing.

We propose Joint MLP/Attention (JoMA) dynamics, a novel mathematical framework to understand the training procedure of multilayer Transformer architectures. This is achieved by integrating out the self-attention layer in Transformers, producing a modified dynamics of MLP layers only. JoMA removes unrealistic assumptions in previous analysis (e.g., lack of residual connection), and predicts that the attention first becomes sparse (to learn salient tokens), then dense (to learn less salient tokens) in the presence of nonlinear activations, while in the linear case, it is consistent with existing works. We leverage JoMA to qualitatively explains how tokens are combined to form hierarchies in multilayer Transformers, when the input tokens are generated by a latent hierarchical generative model. Experiments on models trained from real-world dataset (Wikitext2/Wikitext103) and various pre- trained models (OPT, Pythia) verify our theoretical findings. The code is at https://github.com/facebookresearch/luckmatters/tree/yuandong3.

Deploying Large Language Models (LLMs) in streaming applications such as multi-round dialogue, where long interactions are expected, is urgently needed but poses two major challenges. Firstly, during the decoding stage, caching previous tokens' Key and Value states (KV) consumes extensive memory. Secondly, popular LLMs cannot generalize to longer texts than the training sequence length. Window attention, where only the most recent KVs are cached, is a natural approach --- but we show that it fails when the text length surpasses the cache size. We observe an interesting phenomenon, namely attention sink, that keeping the KV of initial tokens will largely recover the performance of window attention. In this paper, we first demonstrate that the emergence of attention sink is due to the strong attention scores towards initial tokens as a ``sink'' even if they are not semantically important. Based on the above analysis, we introduce StreamingLLM, an efficient framework that enables LLMs trained with a finite length attention window to generalize to infinite sequence length without any fine-tuning. We show that StreamingLLM can enable Llama-2, MPT, Falcon, and Pythia to perform stable and efficient language modeling with up to 4 million tokens and more. In addition, we discover that adding a placeholder token as a dedicated attention sink during pre-training can further improve streaming deployment. In streaming settings, StreamingLLM outperforms the sliding window recomputation baseline by up to 22.2$\times$ speedup. Code and datasets are provided at https://github.com/mit-han-lab/streaming-llm.

We propose Reinforcement Learning from Contrastive Distillation (RLCD), a method for aligning language models to follow principles expressed in natural language (e.g., to be more harmless) without using human feedback. RLCD creates preference pairs from two contrasting model outputs, one using a positive prompt designed to encourage following the given principles, and one using a negative prompt designed to encourage violating them. Using two different prompts causes model outputs to be more differentiated on average, resulting in cleaner preference labels in the absence of human annotations. We then use the preference pairs to train a preference model, which is in turn used to improve a base unaligned language model via reinforcement learning. Empirically, RLCD outperforms RLAIF (Bai et al., 2022b) and context distillation (Huang et al., 2022) baselines across three diverse alignment tasks—harmlessness, helpfulness, and story outline generation—and when using both 7B and 30B model scales for simulating preference data

This paper tackles the memory hurdle of of processing long context sequences in Large Language Models (LLMs), by presenting a novel approach, Dropping In Convolutions for **Lo**ng **Co**ntext **Co**mpression (**LoCoCo**). LoCoCo employs only a fixed-size Key-Value (KV) cache, and can enhance efficiency in both inference and fine-tuning stages. Diverging from prior methods that selectively drop KV pairs based on heuristics, LoCoCo leverages a data-driven adaptive fusion technique, blending previous KV pairs with incoming tokens to minimize the loss of contextual information and ensure accurate attention modeling. This token integration is achieved through injecting one-dimensional convolutional kernels that dynamically calculate mixing weights for each KV cache slot. Designed for broad compatibility with existing LLM frameworks, LoCoCo allows for straightforward "drop-in" integration without needing architectural modifications, while incurring minimal tuning overhead. Experiments demonstrate that LoCoCo maintains consistently outstanding performance across various context lengths and can achieve a high context compression rate during both inference and fine-tuning phases. During inference, we successfully compressed up to $3482$ tokens into a $128$-size KV cache, while retaining comparable performance to the full sequence - an accuracy improvement of up to $0.2791$ compared to baselines at the same cache size. During post-training tuning, we also effectively extended the context length from 4K to 32K using a KV cache of fixed size 512, achieving performance similar to fine-tuning with entire sequences.

Deep generative models like GAN and VAE have shown impressive results in generating unconstrained objects like images. However, many design settings arising in industrial design, material science, computer graphics and more require that the generated objects satisfy hard combinatorial constraints or meet objectives in addition to modeling a data distribution. To address this, we propose GenCO, a generative framework that guarantees constraint satisfaction throughout training by leveraging differentiable combinatorial solvers to enforce feasibility. GenCO imposes the generative loss on provably feasible solutions rather than intermediate soft solutions, meaning that the deep generative network can focus on ensuring the generated objects match the data distribution without having to also capture feasibility. This shift enables practitioners to enforce hard constraints on the generated outputs during end-to-end training, enabling assessments of their feasibility and introducing additional combinatorial loss components to deep generative training. We demonstrate the effectiveness of our approach on a variety of generative combinatorial tasks, including game level generation, map creation for path planning, and photonic device design, consistently demonstrating its capability to yield diverse, high-quality solutions that verifiably adhere to user-specified combinatorial properties.

Mixed integer linear programs (MILP) are flexible and powerful tools for modeling and solving many difficult real-world combinatorial optimization problems. In this paper, we propose a novel machine learning (ML)-based framework ConPaS that learns to predict solutions to MILPs with contrastive learning. For training, we collect high-quality solutions as positive samples. We also collect low-quality or infeasible solutions as negative samples using novel optimization-based or sampling approaches. We then learn to make discriminative predictions by contrasting the positive and negative samples. During testing, we predict and fix the assignments for a subset of integer variables and then solve the resulting reduced MILP to find high-quality solutions. Empirically, ConPaS achieves state-of-the-art results compared to other ML-based approaches in terms of the quality of and the speed at which solutions are found.

This paper addresses the growing need for efficient large language models (LLMs) on mobile devices, driven by increasing cloud costs and latency concerns. We focus on designing top-quality LLMs with fewer than a billion parameters, a practical choice for mobile deployment. Contrary to prevailing belief emphasizing the pivotal role of data and parameter quantity in determining model quality, our investigation underscores the significance of model architecture for sub-billion scale LLMs. Leveraging deep and thin architectures, coupled with embedding sharing and grouped-query attention mechanisms, we establish a strong baseline network denoted as MobileLLM, which attains a remarkable 2.7%/4.3% accuracy boost over preceding 125M/350M state-of-the-art models. Additionally, we propose an immediate block-wise weight-sharing approach with no increase in model size and only marginal latency overhead. The resultant models, denoted as MobileLLM-LS, demonstrate a further accuracy enhancement of 0.7%/0.8% than MobileLLM 125M/350M. Moreover, MobileLLM model family shows significant improvements compared to previous sub-billion models on chat benchmarks, and demonstrates close correctness to LLaMA-v2 7B in API calling tasks, highlighting the capability of small models for common on-device use cases.

Planning has been part of the core pursuit for artificial intelligence since its conception, but earlier AI agents mostly focused on constrained settings because many of the cognitive substrates necessary for human-level planning have been lacking. Recently, language agents powered by large language models (LLMs) have shown interesting capabilities such as tool use and reasoning. Are these language agents capable of planning in more complex settings that are out of the reach of prior AI agents? To advance this investigation, we propose TravelPlanner, a new planning benchmark that focuses on travel planning, a common real-world planning scenario. It provides a rich sandbox environment, various tools for accessing nearly four million data records, and 1,225 meticulously curated planning intents and reference plans. Comprehensive evaluations show that the current language agents are not yet capable of handling such complex planning tasks—even GPT-4 only achieves a success rate of 0.6%. Language agents struggle to stay on task, use the right tools to collect information, or keep track of multiple constraints. However, we note that the mere possibility for language agents to tackle such a complex problem is in itself non-trivial progress. TravelPlanner provides a challenging yet meaningful testbed for future language agents.

Training Large Language Models (LLMs) presents significant memory challenges, predominantly due to the growing size of weights and optimizer states. Common memory-reduction approaches, such as low-rank adaptation (LoRA), add a trainable low-rank matrix to the frozen pre-trained weight in each layer, reducing trainable parameters and optimizer states. However, such approaches typically underperform training with full-rank weights in both pre-training and fine-tuning stages since they limit the parameter search to a low-rank subspace and alter the training dynamics, and further, may require full-rank warm start. In this work, we propose Gradient Low-Rank Projection (GaLore), a training strategy that allows full-parameter learning but is more memory-efficient than common low-rank adaptation methods such as LoRA. Our approach reduces memory usage by up to 65.5% in optimizer states while maintaining both efficiency and performance for pre-training on LLaMA 1B and 7B architectures with C4 dataset with up to 19.7B tokens, and on fine-tuning RoBERTa on GLUE tasks. Our 8-bit GaLore further reduces optimizer memory by up to 82.5% and total training memory by 63.3%, compared to a BF16 baseline. Notably, we demonstrate, for the first time, the feasibility of pre-training a 7B model on consumer GPUs with 24GB memory (e.g., NVIDIA RTX 4090) without model parallel, checkpointing, or offloading strategies.

Conventional wisdom suggests that pre-training Vision Transformers (ViT) improves downstream performance by learning useful representations. Is this actually true? We investigate this question and find that the features and representations learned during pre-training are not essential. Surprisingly, using only the attention patterns from pre-training (i.e., guiding how information flows between tokens) is sufficient for models to learn high quality features from scratch and achieve comparable downstream performance. We show this by introducing a simple method called attention transfer, where only the attention patterns from a pre-trained teacher ViT are transferred to a student, either by copying or distilling the attention maps. Since attention transfer lets the student learn its own features, ensembling it with a fine-tuned teacher also further improves accuracy on ImageNet. We systematically study various aspects of our findings on the sufficiency of attention maps, including distribution shift settings where they underperform fine-tuning. We hope our exploration provides a better understanding of what pre-training accomplishes and leads to a useful alternative to the standard practice of fine-tuning.

Auto-regressive large language models (LLMs) show impressive capacities to solve many complex reasoning tasks while struggling with some simple logical reasoning tasks such as inverse search: when trained on ''$A \to B$'' (e.g., *Tom is the parent of John*), LLM fails to directly conclude ''$B \gets A$'' (e.g., *John is the child of Tom*) during inference even if the two sentences are semantically identical, which is known as the ''reversal curse''. In this paper, we theoretically analyze the reversal curse via the training dynamics of (stochastic) gradient descent for two auto-regressive models: (1) a bilinear model that can be viewed as a simplification of a one-layer transformer; (2) one-layer transformers under certain assumptions. Our analysis reveals that for both models, the reversal curse is a consequence of the (effective) model weights *asymmetry*, i.e., the increase of weights from a token $A$ to token $B$ during training does not necessarily cause the increase of the weights from $B$ to $A$, which is caused by the training dynamics under certain choice of loss function and the optimization space of model parameters. Moreover, our analysis can be naturally applied to other logical reasoning tasks such as chain-of-thought (COT), which provides a new perspective different from previous work that focuses on expressivity. Finally, we conduct experiments to validate our theory on multi-layer transformers under different settings. Our code is available at [https://github.com/marlo-z/reversal_curse_analysis/](https://github.com/marlo-z/reversal_curse_analysis/).

We propose the Detailed Outline Control (DOC) framework for improving long-range plot coherence when automatically generating several-thousand-word-long stories.DOC consists of two complementary components: a detailed outliner and a detailed controller.The detailed outliner creates a more detailed, hierarchically structured outline, shifting creative burden from the main drafting procedure to the planning stage.The detailed controller ensures the more detailed outline is still respected during generation by controlling story passages to align with outline details.In human evaluations of automatically generated stories, DOC substantially outperforms a strong Re 3 baseline (Yang et al., 2022) on plot coherence (22.5% absolute gain), outline relevance (28.2%), and interestingness (20.7%).Humans also judged DOC to be much more controllable in an interactive generation setting. Daisy is a kind-hearted old woman.She has cancer.Bill is her husband.2. Lisa is Daisy's daughter. Structured Prompt For DraftingDaisy is diagnosed with cancer.Lisa is trying to find a viable treatment.Lisa has been stressed out lately, and Daisy expresses her concern.Lisa tirelessly continues her research.Lisa finally finds a cure.Setting: Lisa's laboratory.Lisa looked back at Daisy, her eyes clear and full of determination.

The aggressive performance optimizations in modern microprocessors can result in security vulnerabilities. For example, timing-based attacks in processor caches can steal secret keys or break randomization. So far, finding cache-timing vulnerabilities is mostly performed by human experts, which is inefficient and laborious. There is a need for automatic tools that can explore vulnerabilities given that unreported vulnerabilities leave the systems at risk.In this paper, we propose AutoCAT, an automated exploration framework that finds cache timing-channel attack sequences using reinforcement learning (RL). Specifically, AutoCAT formulates the cache timing-channel attack as a guessing game between an attack program and a victim program holding a secret. This guessing game can thus be solved via modern deep RL techniques. AutoCAT can explore attacks in various cache configurations without knowing design details and under different attack and victim program configurations. AutoCAT can also find attacks to bypass certain detection and defense mechanisms. In particular, AutoCAT discovered StealthyStreamline, a new attack that is able to bypass performance counter-based detection and has up to a 71% higher information leakage rate than the state-of-the-art LRU-based attacks on real processors. AutoCAT is the first of its kind in using RL for crafting microarchitectural timing-channel attack sequences and can accelerate cache timing-channel exploration for secure microprocessor designs.

Optimization problems with nonlinear cost functions and combinatorial constraints appear in many real-world applications but remain challenging to solve efficiently compared to their linear counterparts. To bridge this gap, we propose $\textbf{\emph{\texttt{SurCo}}}$ that learns linear $\underline{\text{Sur}}$rogate costs which can be used in existing $\underline{\text{Co}}$mbinatorial solvers to output good solutions to the original nonlinear combinatorial optimization problem. The surrogate costs are learned end-to-end with nonlinear loss by differentiating through the linear surrogate solver, combining the flexibility of gradient-based methods with the structure of linear combinatorial optimization. We propose three $\texttt{SurCo}$ variants: $\texttt{SurCo}-\texttt{zero}$ for individual nonlinear problems, $\texttt{SurCo}-\texttt{prior}$ for problem distributions, and $\texttt{SurCo}-\texttt{hybrid}$ to combine both distribution and problem-specific information. We give theoretical intuition motivating $\texttt{SurCo}$, and evaluate it empirically. Experiments show that $\texttt{SurCo}$ finds better solutions faster than state-of-the-art and domain expert approaches in real-world optimization problems such as embedding table sharding, inverse photonic design, and nonlinear route planning.

Integer Linear Programs (ILPs) are powerful tools for modeling and solving a large number of combinatorial optimization problems. Recently, it has been shown that Large Neighborhood Search (LNS), as a heuristic algorithm, can find high-quality solutions to ILPs faster than Branch and Bound. However, how to find the right heuristics to maximize the performance of LNS remains an open problem. In this paper, we propose a novel approach, CL-LNS, that delivers state-of-the-art anytime performance on several ILP benchmarks measured by metrics including the primal gap, the primal integral, survival rates and the best performing rate. Specifically, CL-LNS collects positive and negative solution samples from an expert heuristic that is slow to compute and learns a more efficient one with contrastive learning. We use graph attention networks and a richer set of features to further improve its performance.

Finding the optimal pass sequence of compilation can lead to a significant reduction in program size. Prior works on compilation pass ordering have two major drawbacks. They either require an excessive budget (in terms of the number of compilation passes) at compile time or fail to generalize to unseen programs. In this work, instead of predicting passes sequentially, we directly learn a policy on the pass sequence space, which outperforms the default -Oz flag by an average of 4.5% over a large collection (4683) of unseen code repositories from diverse domains across 14 datasets. To achieve this, we first identify a small set (termed coreset) of pass sequences that generally optimize the size of most programs. Then, a policy is learned to pick the optimal sequences by predicting the normalized values of the pass sequences in the coreset. Our results demonstrate that existing human-designed compiler passes can be improved with a simple yet effective technique that leverages pass sequence space which contains dense rewards, while approaches operating on the individual pass space may suffer from issues of sparse reward, and do not generalize well to held-out programs from different domains. Website: https://rlcompopt.github.io.

Large language models (LLMs) with hundreds of billions of parameters have sparked a new wave of exciting AI applications. However, they are computationally expensive at inference time. Sparsity is a natural approach to reduce this cost, but existing methods either require costly retraining, have to forgo LLM's in-context learning ability, or do not yield wall-clock time speedup on modern hardware. We hypothesize that contextual sparsity, which are small, input-dependent sets of attention heads and MLP parameters that yield approximately the same output as the dense model for a given input, can address these issues. We show that contextual sparsity exists, that it can be accurately predicted, and that we can exploit it to speed up LLM inference in wall-clock time without compromising LLM's quality or in-context learning ability. Based on these insights, we propose DejaVu, a system that uses a low-cost algorithm to predict contextual sparsity on the fly given inputs to each layer, along with an asynchronous and hardware-aware implementation that speeds up LLM inference. We validate that DejaVu can reduce the inference latency of OPT-175B by over 2$\times$ compared to the state-of-the-art FasterTransformer, and over 6$\times$ compared to the widely used Hugging Face implementation, without compromising model quality. The code is available at https://github.com/FMInference/DejaVu.

Transformer architecture has shown impressive performance in multiple research domains and has become the backbone of many neural network models. However, there is limited understanding on how it works. In particular, with a simple predictive loss, how the representation emerges from the gradient \emph{training dynamics} remains a mystery. In this paper, for 1-layer transformer with one self-attention layer plus one decoder layer, we analyze its SGD training dynamics for the task of next token prediction in a mathematically rigorous manner. We open the black box of the dynamic process of how the self-attention layer combines input tokens, and reveal the nature of underlying inductive bias. More specifically, with the assumption (a) no positional encoding, (b) long input sequence, and (c) the decoder layer learns faster than the self-attention layer, we prove that self-attention acts as a \emph{discriminative scanning algorithm}: starting from uniform attention, it gradually attends more to distinct key tokens for a specific next token to be predicted, and pays less attention to common key tokens that occur across different next tokens. Among distinct tokens, it progressively drops attention weights, following the order of low to high co-occurrence between the key and the query token in the training set. Interestingly, this procedure does not lead to winner-takes-all, but stops due to a \emph{phase transition} that is controllable by the learning rate of the decoder layer, leaving (almost) fixed token combination. We verify this \textbf{\emph{scan and snap}} dynamics on synthetic and real-world data (WikiText-103).

Large Language Models (LLMs), despite their recent impressive accomplishments, are notably cost-prohibitive to deploy, particularly for applications involving long-content generation, such as dialogue systems and story writing. Often, a large amount of transient state information, referred to as the $\mathsf{KV}$ $\mathsf{cache}$, is stored in GPU memory in addition to model parameters, scaling linearly with the sequence length and batch size. In this paper, we introduce a novel approach for implementing the $\mathsf{KV}$ $\mathsf{cache}$ which significantly reduces its memory footprint. Our approach is based on the noteworthy observation that a small portion of tokens contributes most of the value when computing attention scores. We call these tokens Heavy Hitters ($\mathsf{H_2}$). Through a comprehensive investigation, we find that ($i$) the emergence of $\mathsf{H_2}$ is natural and strongly correlates with the frequent co-occurrence of tokens in the text, and ($ii$) removing them results in significant performance degradation. Based on these insights, we propose Heavy Hitter Oracle ($\mathsf{H_2O}$), a $\mathsf{KV}$ $\mathsf{cache}$ eviction policy that dynamically retains a balance of recent and $\mathsf{H_2}$ tokens. We formulate the $\mathsf{KV}$ $\mathsf{cache}$ eviction as a dynamic submodular problem and prove (under mild assumptions) a theoretical guarantee for our novel eviction algorithm which could help guide future work. We validate the accuracy of our algorithm with OPT, LLaMA, and GPT-NeoX across a wide range of tasks. Our implementation of $\mathsf{H_2O}$ with 20\% heavy hitters improves the throughput over three leading inference systems DeepSpeed Zero-Inference, Hugging Face Accelerate, and FlexGen by up to $29\times$, $29\times$, and $3\times$ on OPT-6.7B and OPT-30B. With the same batch size, $\mathsf{H_2O}$ can reduce the latency by up to $1.9\times$.

Recent works in learning-integrated optimization have shown promise in settings where the optimization problem is only partially observed or where general-purpose optimizers perform poorly without expert tuning. By learning an optimizer $\mathbf{g}$ to tackle these challenging problems with $f$ as the objective, the optimization process can be substantially accelerated by leveraging past experience. The optimizer can be trained with supervision from known optimal solutions or implicitly by optimizing the compound function $f\circ \mathbf{g}$. The implicit approach may not require optimal solutions as labels and is capable of handling problem uncertainty; however, it is slow to train and deploy due to frequent calls to optimizer $\mathbf{g}$ during both training and testing. The training is further challenged by sparse gradients of $\mathbf{g}$, especially for combinatorial solvers. To address these challenges, we propose using a smooth and learnable **Landscape Surrogate** $\mathcal{M}$ as a replacement for $f\circ \mathbf{g}$. This surrogate, learnable by neural networks, can be computed faster than the solver $\mathbf{g}$, provides dense and smooth gradients during training, can generalize to unseen optimization problems, and is efficiently learned via alternating optimization. We test our approach on both synthetic problems, including shortest path and multidimensional knapsack, and real-world problems such as portfolio optimization, achieving comparable or superior objective values compared to state-of-the-art baselines while reducing the number of calls to $\mathbf{g}$. Notably, our approach outperforms existing methods for computationally expensive high-dimensional problems.

Transformers have achieved great success in several domains, including Natural Language Processing and Computer Vision. However, their application to real-world graphs is less explored, mainly due to its high computation cost and its poor generalizability caused by the lack of enough training data in the graph domain. To fill in this gap, we propose a scalable Transformer-like dynamic graph learning method named Dynamic Graph Transformer (DyFormer) with spatial-temporal encoding to effectively learn graph topology and capture implicit links. To achieve efficient and scalable training, we propose temporal-union graph structure and its associated subgraph-based node sampling strategy. To improve the generalization ability, we introduce two complementary self-supervised, pre-training tasks and show that jointly optimizing the two pre-training tasks results in a smaller Bayesian error rate via an information- theoretic analysis. Extensive experiments on the real- world datasets illustrate that DyFormer achieves a consistent 1% ~ 3% AUC gain (averaged over all time steps) compared with baselines on all benchmarks. [Code]

This paper presents the design, implementation, evaluation, and deployment of Meta's production network migration system. We first introduce the network migration problem for large-scale production datacenter networks (DCNs). A network migration task at Meta touches as many as hundreds of switches and tens of thousands of circuits per datacenter (DC), and involves physical deployment work on site that can last months. We describe real-world migration challenges, covering complex and evolving DCN architectures and operational constraints. We mathematically formalize the problem of generating efficient and safe migration plans, and exploit the inherent symmetry and locality of DCN topologies to prune the search space. We design an ordering-agnostic compact topology representation to eliminate redundant satisfiability checking, and apply the A* algorithm with a domain-specific priority function to find the optimal plan. Evaluation results on a range of production migration cases show that Klotski reduces the time to find optimal migration plans by up to 381× compared to prior solutions. We hope by introducing the problem and sharing our deployment experience, this work can provide a useful context for network migration in the real world and inspire future research.

The high computation cost is one of the key bottlenecks for adopting deep neural networks (DNNs) in different hardware. When client data are sensitive, privacy-preserving DNN evaluation method, such as homomorphic encryptions (HE), shows even more computation cost. Prior works employed weight repetition in quantized neural networks to save the computation of convolutions by memorizing or arithmetic factorization. However, such methods fail to fully exploit the exponential search space from factorizing and reusing computation. We propose Q-gym, a DNN framework consisting of two components. First, we propose a compiler, which leverages equality saturation to generate computation expressions for convolutional layers with a significant reduction in the number of operations. Second, we integrate the computation expressions with various parallelization methods to accelerate DNN inference on different hardware. We also employ the efficient expressions to accelerate DNN inference under HE.

Long Short-Term Memory (LSTM) and Transformers are two popular neural architectures used for natural language processing tasks. Theoretical results show that both are Turing-complete and can represent any context-free language (CFL).In practice, it is often observed that Transformer models have better representation power than LSTM. But the reason is barely understood. We study such practical differences between LSTM and Transformer and propose an explanation based on their latent space decomposition patterns. To achieve this goal, we introduce an oracle training paradigm, which forces the decomposition of the latent representation of LSTMand the Transformer and supervises with the transitions of the Pushdown Automaton (PDA) of the corresponding CFL. With the forced decomposition, we show that the performance upper bounds of LSTM and Transformer in learning CFL are close: both of them can simulate a stack and perform stack operation along with state transitions. However, the absence of forced decomposition leads to the failure of LSTM models to capture the stack and stack operations, while having a marginal impact on the Transformer model. Lastly, we connect the experiment on the prototypical PDA to a real-world parsing task to re-verify the conclusions

Policy-based methods with function approximation are widely used for solving two-player zero-sum games with large state and/or action spaces. However, it remains elusive how to obtain optimization and statistical guarantees for such algorithms. We present a new policy optimization algorithm with function approximation and prove that under standard regularity conditions on the Markov game and the function approximation class, our algorithm finds a near-optimal policy within a polynomial number of samples and iterations. To our knowledge, this is the first provably efficient policy optimization algorithm with function approximation that solves two-player zero-sum Markov games.

Interest in applying Artificial Intelligence (AI) techniques to compiler optimizations is increasing rapidly, but compiler research has a high entry barrier. Unlike in other domains, compiler and AI researchers do not have access to the datasets and frameworks that enable fast iteration and development of ideas, and getting started requires a significant engineering investment. What is needed is an easy, reusable experimental infrastructure for real world compiler optimization tasks that can serve as a common benchmark for comparing techniques, and as a platform to accelerate progress in the field.We introduce CompilerGym, a set of environments for real world compiler optimization tasks, and a toolkit for exposing new optimization tasks to compiler researchers. CompilerGym enables anyone to experiment on production compiler optimization problems through an easy-to-use package, regardless of their experience with compilers. We build upon the popular OpenAI Gym interface enabling researchers to interact with compilers using Python and a familiar API.We describe the CompilerGym architecture and implementation, characterize the optimization spaces and computational efficiencies of three included compiler environments, and provide extensive empirical evaluations. Compared to prior works, CompilerGym offers larger datasets and optimization spaces, is 27× more computationally efficient, is fault-tolerant, and capable of detecting reproducibility bugs in the underlying compilers.In making it easy for anyone to experiment with compilers - irrespective of their background - we aim to accelerate progress in the AI and compiler research domains.

Many recent self-supervised frameworks for visual representation learning are based on certain forms of Siamese networks. Such networks are conceptually symmetric with two parallel encoders, but often practically asymmetric as numerous mechanisms are devised to break the symmetry. In this work, we conduct a formal study on the importance of asymmetry by explicitly distinguishing the two encoders within the network - one produces source encodings and the other targets. Our key insight is keeping a relatively lower variance in target than source generally benefits learning. This is empirically justified by our results from five case studies covering different variance-oriented designs, and is aligned with our preliminary theoretical analysis on the baseline. Moreover, we find the improvements from asymmetric designs generalize well to longer training schedules, multiple other frameworks and newer backbones. Finally, the combined effect of several asymmetric designs achieves a state-of-the-art accuracy on ImageNet linear probing and competitive results on downstream transfer. We hope our exploration will inspire more research in exploiting asymmetry for Siamese representation learning.

We consider the problem of automatically generating longer stories of over two thousand words.Compared to prior work on shorter stories, long-range plot coherence and relevance are more central challenges here.We propose the Recursive Reprompting and Revision framework (Re 3 ) to address these challenges by (a) prompting a general-purpose language model to construct a structured overarching plan, and (b) generating story passages by repeatedly injecting contextual information from both the plan and current story state into a language model prompt.We then revise by (c) reranking different continuations for plot coherence and premise relevance, and finally (d) editing the best continuation for factual consistency.Compared to similar-length stories generated directly from the same base model, human evaluators judged substantially more of Re 3 's stories as having a coherent overarching plot (by 14% absolute increase), and relevant to the given initial premise (by 20%). Autoregressive Context EditPeyton Turner Peyton Turner is male.Peyton works at a restaurant. Inferred FactsShe knew Peyton was probably

The ability to separate signal from noise, and reason with clean abstractions, is critical to intelligence. With this ability, humans can efficiently perform real world tasks without considering all possible nuisance factors. How can artificial agents do the same? What kind of information can agents safely discard as noises? In this work, we categorize information out in the wild into four types based on controllability and relation with reward, and formulate useful information as that which is both controllable and reward-relevant. This framework clarifies the kinds information removed by various prior work on representation learning in reinforcement learning (RL), and leads to our proposed approach of learning a Denoised MDP that explicitly factors out certain noise distractors. Extensive experiments on variants of DeepMind Control Suite and RoboDesk demonstrate superior performance of our denoised world model over using raw observations alone, and over prior works, across policy optimization control tasks as well as the non-control task of joint position regression. Project Page: https://ssnl.github.io/denoised_mdp/ Code: https://github.com/facebookresearch/denoised_mdp/

Embedding learning is an important technique in deep recommendation models to map categorical features to dense vectors. However, the embedding tables often demand an extremely large number of parameters, which become the storage and efficiency bottlenecks. Distributed training solutions have been adopted to partition the embedding tables into multiple devices. However, the embedding tables can easily lead to imbalances if not carefully partitioned. This is a significant design challenge of distributed systems named embedding table sharding, i.e., how we should partition the embedding tables to balance the costs across devices, which is a non-trivial task because 1) it is hard to efficiently and precisely measure the cost, and 2) the partition problem is known to be NP-hard. In this work, we introduce our novel practice in Meta, namely AutoShard, which uses a neural cost model to directly predict the multi-table costs and leverages deep reinforcement learning to solve the partition problem. Experimental results on an open-sourced large-scale synthetic dataset and Meta's production dataset demonstrate the superiority of AutoShard over the heuristics. Moreover, the learned policy of AutoShard can transfer to sharding tasks with various numbers of tables and different ratios of the unseen tables without any fine-tuning. Furthermore, AutoShard can efficiently shard hundreds of tables in seconds. The effectiveness, transferability, and efficiency of AutoShard make it desirable for production use. Our algorithms have been deployed in Meta production environment. A prototype is available at https://github.com/daochenzha/autoshard

We show that Contrastive Learning (CL) under a broad family of loss functions (including InfoNCE) has a unified formulation of coordinate-wise optimization on the network parameter $\vtheta$ and pairwise importance $\alpha$, where the \emph{max player} $\vtheta$ learns representation for contrastiveness, and the \emph{min player} $\alpha$ puts more weights on pairs of distinct samples that share similar representations. The resulting formulation, called \boldmethod{}, unifies not only various existing contrastive losses, which differ by how sample-pair importance $\alpha$ is constructed, but also is able to extrapolate to give novel contrastive losses beyond popular ones, opening a new avenue of contrastive loss design. These novel losses yield comparable (or better) performance on CIFAR10, STL-10 and CIFAR-100 than classic InfoNCE. Furthermore, we also analyze the max player in detail: we prove that with fixed $\alpha$, max player is equivalent to Principal Component Analysis (PCA) for deep linear network, and almost all local minima are global and rank-1, recovering optimal PCA solutions. Finally, we extend our analysis on max player to 2-layer ReLU networks, showing that its fixed points can have higher ranks. Codes are available in https://github.com/facebookresearch/luckmatters/tree/main/ssl/real-dataset.

We study embedding table placement for distributed recommender systems, which aims to partition and place the tables on multiple hardware devices (e.g., GPUs) to balance the computation and communication costs. Although prior work has explored learning-based approaches for the device placement of computational graphs, embedding table placement remains to be a challenging problem because of 1) the operation fusion of embedding tables, and 2) the generalizability requirement on unseen placement tasks with different numbers of tables and/or devices. To this end, we present DreamShard, a reinforcement learning (RL) approach for embedding table placement. DreamShard achieves the reasoning of operation fusion and generalizability with 1) a cost network to directly predict the costs of the fused operation, and 2) a policy network that is efficiently trained on an estimated Markov decision process (MDP) without real GPU execution, where the states and the rewards are estimated with the cost network. Equipped with sum and max representation reductions, the two networks can directly generalize to any unseen tasks with different numbers of tables and/or devices without fine-tuning. Extensive experiments show that DreamShard substantially outperforms the existing human expert and RNN-based strategies with up to 19% speedup over the strongest baseline on large-scale synthetic tables and our production tables. The code is available.

Neural Architecture Search (NAS) has emerged as a promising technique for automatic neural network design. However, existing MCTS based NAS approaches often utilize manually designed action space, which is not directly related to the performance metric to be optimized (e.g., accuracy), leading to sample-inefficient explorations of architectures. To improve the sample efficiency, this paper proposes Latent Action Neural Architecture Search (LaNAS), which learns actions to recursively partition the search space into good or bad regions that contain networks with similar performance metrics. During the search phase, as different action sequences lead to regions with different performance, the search efficiency can be significantly improved by biasing towards the good regions. On three NAS tasks, empirical results demonstrate that LaNAS is at least an order more sample efficient than baseline methods including evolutionary algorithms, Bayesian optimizations, and random search. When applied in practice, both one-shot and regular LaNAS consistently outperform existing results. Particularly, LaNAS achieves 99.0 percent accuracy on CIFAR-10 and 80.8 percent top1 accuracy at 600 MFLOPS on ImageNet in only 800 samples, significantly outperforming AmoebaNet with 33× fewer samples. Our code is publicly available at https://github.com/facebookresearch/LaMCTS.

Adversarial examples have appeared as a ubiquitous property of machine learning models where bounded adversarial perturbation could mislead the models to make arbitrarily incorrect predictions. Such examples provide a way to assess the robustness of machine learning models as well as a proxy for understanding the model training process. Extensive studies try to explain the existence of adversarial examples and provide ways to improve model robustness (e.g. adversarial training). While they mostly focus on models trained on datasets with predefined labels, we leverage the teacher-student framework and assume a teacher model, or \emph{oracle}, to provide the labels for given instances. We extend \citet{tian2019student} in the case of low-rank input data and show that \emph{student specialization} (trained student neuron is highly correlated with certain teacher neuron at the same layer) still happens within the input subspace, but the teacher and student nodes could \emph{differ wildly} out of the data subspace, which we conjecture leads to adversarial examples. Extensive experiments show that student specialization correlates strongly with model robustness in different scenarios, including student trained via standard training, adversarial training, confidence-calibrated adversarial training, and training with robust feature dataset. Our studies could shed light on the future exploration about adversarial examples, and enhancing model robustness via principled data augmentation.

Neural Architecture Search (NAS) yields state-of-the-art neural networks that outperform their best manually-designed counterparts. However, previous NAS methods search for architectures under one set of training hyper-parameters (i.e., a training recipe), overlooking superior architecture-recipe combinations. To address this, we present Neural Architecture-Recipe Search (NARS) to search both (a) architectures and (b) their corresponding training recipes, simultaneously. NARS utilizes an accuracy predictor that scores architecture and training recipes jointly, guiding both sample selection and ranking. Furthermore, to compensate for the enlarged search space, we leverage "free" architecture statistics (e.g., FLOP count) to pretrain the predictor, significantly improving its sample efficiency and prediction reliability. After training the predictor via constrained iterative optimization, we run fast evolutionary searches in just CPU minutes to generate architecturerecipe pairs for a variety of resource constraints, called FBNetV3. FBNetV3 makes up a family of state-of-the-art compact neural networks that outperform both automatically and manually-designed competitors. For example, FB-NetV3 matches both EfficientNet and ResNeSt accuracy on ImageNet with up to 2.0× and 7.1 × fewer FLOPs, respectively. Furthermore, FBNetV3 yields significant performance gains for downstream object detection tasks, improving mAP despite 18% fewer FLOPs and 34% fewer parameters than EfficientNet-based equivalents.

Differential Neural Architecture Search (NAS) requires all layer choices to be held in memory simultaneously; this limits the size of both search space and final architecture. In contrast, Probabilistic NAS, such as PARSEC, learns a distribution over high-performing architectures, and uses only as much memory as needed to train a single model. Nevertheless, it needs to sample many architectures, making it computationally expensive for searching in an extensive space. To solve these problems, we propose a sampling method adaptive to the distribution entropy, drawing more samples to encourage explorations at the beginning, and reducing samples as learning proceeds. Furthermore, to search fast in the multivariate space, we propose a coarse-to-fine strategy by using a factorized distribution at the beginning which can reduce the number of architecture parameters by over an order of magnitude. We call this method Fast Probabilistic NAS (FP-NAS). Compared with PARSEC, it can sample 64% fewer architectures and search 2.1× faster. Compared with FBNetV2, FP-NAS is 1.9× - 3.5 ×faster, and the searched models outperform FBNetV2 models on ImageNet. FP-NAS allows us to expand the giant FBNetV2 space to be wider (i.e. larger channel choices) and deeper (i.e. more blocks), while adding Split-Attention block and enabling the search over the number of splits. When searching a model of size 0.4G FLOPS, FP-NAS is 132× faster than EfficientNet, and the searched FP-NAS-L0 model outperforms EfficientNet-B0 by 0.7% accuracy. Without using any architecture surrogate or scaling tricks, we directly search large models up to 1.0G FLOPS. Our FP-NAS-L2 model with simple distillation out-performs BigNAS-XL with advanced inplace distillation by 0.7% accuracy using similar FLOPS.

Task-specific fine-tuning on pre-trained transformers has achieved performance breakthroughs in multiple NLP tasks. Yet, as both computation and parameter size grows linearly with the number of sub-tasks, it is increasingly difficult to adopt such methods to the real world due to unrealistic memory and computation overhead on computing devices. Previous works on fine-tuning focus on reducing the growing parameter size to save storage cost by parameter sharing. However, compared to storage, the constraint of computation is a more critical issue with the fine-tuning models in modern computing environments. In this work, we propose LeTS, a framework that leverages both computation and parameter sharing across multiple tasks. Compared to traditional fine-tuning, LeTS proposes a novel neural architecture that contains a fixed pre-trained transformer model, plus learnable additive components for sub-tasks. The learnable components reuse the intermediate activations in the fixed pre-trained model, decoupling computation dependency. Differentiable neural architecture search is used to determine a task-specific computation sharing scheme, and a novel early stage pruning is applied to additive components for sparsity to achieve parameter sharing. Extensive experiments show that with 1.4% of extra parameters per task, LeTS reduces the computation by 49.5% on GLUE benchmarks with only 0.2% accuracy loss compared to full fine-tuning.

While contrastive approaches of self-supervised learning (SSL) learn representations by minimizing the distance between two augmented views of the same data point (positive pairs) and maximizing views from different data points (negative pairs), recent \emph{non-contrastive} SSL (e.g., BYOL and SimSiam) show remarkable performance {\it without} negative pairs, with an extra learnable predictor and a stop-gradient operation. A fundamental question rises: why they do not collapse into trivial representation? In this paper, we answer this question via a simple theoretical study and propose a novel approach, \ourmethod{}, that \emph{directly} sets the linear predictor based on the statistics of its inputs, rather than trained with gradient update. On ImageNet, it performs comparably with more complex two-layer non-linear predictors that employ BatchNorm and outperforms linear predictor by $2.5%$ in 300-epoch training (and $5%$ in 60-epoch). \ourmethod{} is motivated by our theoretical study of the nonlinear learning dynamics of non-contrastive SSL in simple linear networks. Our study yields conceptual insights into how non-contrastive SSL methods learn, how they avoid representational collapse, and how multiple factors, like predictor networks, stop-gradients, exponential moving averages, and weight decay all come into play. Our simple theory recapitulates the results of real-world ablation studies in both STL-10 and ImageNet. Code is released\footnote{\url{https://github.com/facebookresearch/luckmatters/tree/master/ssl}}.

Efficient evaluation of a network architecture drawn from a large search space remains a key challenge in Neural Architecture Search (NAS). Vanilla NAS evaluates each architecture by training from scratch, which gives the true performance but is extremely time-consuming. Recently, one-shot NAS substantially reduces the computation cost by training only one supernetwork, a.k.a. supernet, to approximate the performance of every architecture in the search space via weight-sharing. However, the performance estimation can be very inaccurate due to the co-adaption among operations. In this paper, we propose few-shot NAS that uses multiple supernetworks, called sub-supernet, each covering different regions of the search space to alleviate the undesired co-adaption. Compared to one-shot NAS, few-shot NAS improves the accuracy of architecture evaluation with a small increase of evaluation cost. With only up to 7 sub-supernets, few-shot NAS establishes new SoTAs: on ImageNet, it finds models that reach 80.5% top-1 accuracy at 600 MB FLOPS and 77.5% top-1 accuracy at 238 MFLOPS; on CIFAR10, it reaches 98.72% top-1 accuracy without using extra data or transfer learning. In Auto-GAN, few-shot NAS outperforms the previously published results by up to 20%. Extensive experiments show that few-shot NAS significantly improves various one-shot methods, including 4 gradient-based and 6 search-based methods on 3 different tasks in NasBench-201 and NasBench1-shot-1.

Program synthesis from input-output (IO) examples has been a long-standing challenge. While recent works demonstrated limited success on domain-specific languages (DSL), it remains highly challenging to apply them to real-world programming languages, such as C. Due to complicated syntax and token variation, there are three major challenges: (1) unlike many DSLs, programs in languages like C need to compile first and are not executed via interpreters; (2) the program search space grows exponentially when the syntax and semantics of the programming language become more complex; and (3) collecting a large-scale dataset of real-world programs is non-trivial. As a first step to address these challenges, we propose LaSynth and show its efficacy in a restricted-C domain (i.e., C code with tens of tokens, with sequential, branching, loop and simple arithmetic operations but no library call). More specifically, LaSynth learns the latent representation to approximate the execution of partially generated programs, even if they are incomplete in syntax (addressing (1)). The learned execution significantly improves the performance of next token prediction over existing approaches, facilitating search (addressing (2)). Finally, once trained with randomly generated ground-truth programs and their IO pairs, LaSynth can synthesize more concise programs that resemble human-written code. Furthermore, retraining our model with these synthesized programs yields better performance with fewer samples for both Karel and C program synthesis, indicating the promise of leveraging the learned program synthesizer to improve the dataset quality for input-output program synthesis (addressing (3)). When evaluating on whether the program execution outputs match the IO pairs, LaSynth achieves 55.2% accuracy on generating simple C code with tens of tokens including loops and branches, outperforming existing approaches without executors by around 20%.

Path planning, the problem of efficiently discovering high-reward trajectories, often requires optimizing a high-dimensional and multimodal reward function. Popular approaches like CEM and CMA-ES greedily focus on promising regions of the search space and may get trapped in local maxima. DOO and VOOT balance exploration and exploitation, but use space partitioning strategies independent of the reward function to be optimized. Recently, LaMCTS empirically learns to partition the search space in a reward-sensitive manner for black-box optimization. In this paper, we develop a novel formal regret analysis for when and why such an adaptive region partitioning scheme works. We also propose a new path planning method LaP3 which improves the function value estimation within each sub-region, and uses a latent representation of the search space. Empirically, LaP3 outperforms existing path planning methods in 2D navigation tasks, especially in the presence of difficult-to-escape local optima, and shows benefits when plugged into the planning components of model-based RL such as PETS. These gains transfer to highly multimodal real-world tasks, where we outperform strong baselines in compiler phase ordering by up to 39% on average across 9 tasks, and in molecular design by up to 0.4 on properties on a 0-1 scale. Code is available at https://github.com/yangkevin2/neurips2021-lap3.

In online reinforcement learning (RL), efficient exploration remains particularly challenging in high-dimensional environments with sparse rewards. In low-dimensional environments, where tabular parameterization is possible, count-based upper confidence bound (UCB) exploration methods achieve minimax near-optimal rates. However, it remains unclear how to efficiently implement UCB in realistic RL tasks that involve non-linear function approximation. To address this, we propose a new exploration approach via maximizing the deviation of the occupancy of the next policy from the explored regions. We add this term as an adaptive regularizer to the standard RL objective to balance exploration vs. exploitation. We pair the new objective with a provably convergent algorithm, giving rise to a new intrinsic reward that adjusts existing bonuses. The proposed intrinsic reward is easy to implement and combine with other existing RL algorithms to conduct exploration. As a proof of concept, we evaluate the new intrinsic reward on tabular examples across a variety of model-based and model-free algorithms, showing improvements over count-only exploration strategies. When tested on navigation and locomotion tasks from MiniGrid and DeepMind Control Suite benchmarks, our approach significantly improves sample efficiency over state-of-the-art methods.

Efficient exploration under sparse rewards remains a key challenge in deep reinforcement learning. Previous exploration methods (e.g., RND) have achieved strong results in multiple hard tasks. However, if there are multiple novel areas to explore, these methods often focus quickly on one without sufficiently trying others (like a depth-wise first search manner). In some scenarios (e.g., four corridor environment in Sec 4.2), we observe they explore in one corridor for long and fail to cover all the states. On the other hand, in theoretical RL, with optimistic initialization and the inverse square root of visitation count as a bonus, it won't suffer from this and explores different novel regions alternatively (like a breadth-first search manner). In this paper, inspired by this, we propose a simple but effective criterion called NovelD by weighting every novel area approximately equally. Our algorithm is very simple but yet shows comparable performance or even outperforms multiple SOTA exploration methods in many hard exploration tasks. Specifically, NovelD solves all the static procedurally-generated tasks in Mini-Grid with just 120M environment steps, without any curriculum learning. In comparison, the previous SOTA only solves 50% of them. NovelD also achieves SOTA on multiple tasks in NetHack, a rogue-like game that contains more challenging procedurally-generated environments. In multiple Atari games (e.g., MonteZuma's Revenge, Venture, Gravitar), NovelD outperforms RND. We analyze NovelD thoroughly in MiniGrid and found that empirically it helps the agent explore the environment more uniformly with a focus on exploring beyond the boundary.

Network planning is critical to the performance, reliability and cost of web services. This problem is typically formulated as an Integer Linear Programming (ILP) problem. Today's practice relies on hand-tuned heuristics from human experts to address the scalability challenge of ILP solvers.

Neural Architecture Search (NAS) has shown great success in automating the design of neural networks, but the prohibitive amount of computations behind current NAS methods requires further investigations in improving the sample efficiency and the network evaluation cost to get better results in a shorter time. In this paper, we present a novel scalable Monte Carlo Tree Search (MCTS) based NAS agent, named AlphaX, to tackle these two aspects. AlphaX improves the search efficiency by adaptively balancing the exploration and exploitation at the state level, and by a Meta-Deep Neural Network (DNN) to predict network accuracies for biasing the search toward a promising region. To amortize the network evaluation cost, AlphaX accelerates MCTS rollouts with a distributed design and reduces the number of epochs in evaluating a network by transfer learning, which is guided with the tree structure in MCTS. In 12 GPU days and 1000 samples, AlphaX found an architecture that reaches 97.84% top-1 accuracy on CIFAR-10, and 75.5% top-1 accuracy on ImageNet, exceeding SOTA NAS methods in both the accuracy and sampling efficiency. Particularly, we also evaluate AlphaX on NASBench-101, a large scale NAS dataset; AlphaX is 3x and 2.8x more sample efficient than Random Search and Regularized Evolution in finding the global optimum. Finally, we show the searched architecture improves a variety of vision applications from Neural Style Transfer, to Image Captioning and Object Detection.

Differentiable Neural Architecture Search (DNAS) has demonstrated great success in designing state-of-the-art, efficient neural networks. However, DARTS-based DNAS's search space is small when compared to other search methods', since all candidate network layers must be explicitly instantiated in memory. To address this bottleneck, we propose a memory and computationally efficient DNAS variant: DMaskingNAS. This algorithm expands the search space by up to 10^14x over conventional DNAS, supporting searches over spatial and channel dimensions that are otherwise prohibitively expensive: input resolution and number of filters. We propose a masking mechanism for feature map reuse, so that memory and computational costs stay nearly constant as the search space expands. Furthermore, we employ effective shape propagation to maximize per-FLOP or per-parameter accuracy. The searched FBNetV2s yield state-of-the-art performance when compared with all previous architectures. With up to 421x less search cost, DMaskingNAS finds models with 0.9% higher accuracy, 15% fewer FLOPs than MobileNetV3-Small; and with similar accuracy but 20% fewer FLOPs than Efficient-B0. Furthermore, our FBNetV2 outperforms MobileNetV3 by 2.6% in accuracy, with equivalent model size. FBNetV2 models are open-sourced at https://github.com/facebookresearch/mobile-vision.

We consider a deep ReLU / Leaky ReLU student network trained from the output of a fixed teacher network of the same depth, with Stochastic Gradient Descent (SGD). The student network is \emph{over-realized}: at each layer $l$, the number $n_l$ of student nodes is more than that ($m_l$) of teacher. Under mild conditions on dataset and teacher network, we prove that when the gradient is small at every data sample, each teacher node is \emph{specialized} by at least one student node \emph{at the lowest layer}. For two-layer network, such specialization can be achieved by training on any dataset of \emph{polynomial} size $\mathcal{O}( K^{5/2} d^3 \epsilon^{-1})$. until the gradient magnitude drops to $\mathcal{O}(\epsilon/K^{3/2}\sqrt{d})$. Here $d$ is the input dimension, $K = m_1 + n_1$ is the total number of neurons in the lowest layer of teacher and student. Note that we require a specific form of data augmentation and the sample complexity includes the additional data generated from augmentation. To our best knowledge, we are the first to give polynomial sample complexity for student specialization of training two-layer (Leaky) ReLU networks with finite depth and width in teacher-student setting, and finite complexity for the lowest layer specialization in multi-layer case, without parametric assumption of the input (like Gaussian). Our theory suggests that teacher nodes with large fan-out weights get specialized first when the gradient is still large, while others are specialized with small gradient, which suggests inductive bias in training. This shapes the stage of training as empirically observed in multiple previous works. Experiments on synthetic and CIFAR10 verify our findings. The code is released in \url{https://github.com/facebookresearch/luckmatters}.

Click-Through Rate (CTR) prediction is one of the most important machine learning tasks in recommender systems, driving personalized experience for billions of consumers. Neural architecture search (NAS), as an emerging field, has demonstrated its capabilities in discovering powerful neural network architectures, which motivates us to explore its potential for CTR predictions. Due to 1) diverse unstructured feature interactions, 2) heterogeneous feature space, and 3) high data volume and intrinsic data randomness, it is challenging to construct, search, and compare different architectures effectively for recommendation models. To address these challenges, we propose an automated interaction architecture discovering framework for CTR prediction named AutoCTR. Via modularizing simple yet representative interactions as virtual building blocks and wiring them into a space of direct acyclic graphs, AutoCTR performs evolutionary architecture exploration with learning-to-rank guidance at the architecture level and achieves acceleration using low-fidelity model. Empirical analysis demonstrates the effectiveness of AutoCTR on different datasets comparing to human-crafted architectures. The discovered architecture also enjoys generalizability and transferability among different datasets.

To learn good joint policies for multi-agent collaboration with incomplete information remains a fundamental challenge. While for two-player zero-sum games, coordinate-ascent approaches (optimizing one agent's policy at a time, e.g., self-play) work with guarantees, in multi-agent cooperative setting they often converge to sub-optimal Nash equilibrium. On the other hand, directly modeling joint policy changes in incomplete information game is nontrivial due to complicated interplay of policies (e.g., upstream updates affect downstream state reachability). In this paper, we show global changes of game values can be decomposed to policy changes localized at each information set, with a novel term named \emph{policy-change density}. Based on this, we propose \emph{Joint Policy Search} (JPS) that iteratively improves joint policies of collaborative agents in incomplete information games, without re-evaluating the entire game. On multiple collaborative tabular games, JPS is proven to never worsen performance and can improve solutions provided by unilateral approaches (e.g, CFR), outperforming algorithms designed for collaborative policy learning (e.g. BAD). Furthermore, for real-world game whose states are too many to enumerate, \ours{} has an online form that naturally links with gradient updates. We test it to Contract Bridge, a 4-player imperfect-information game where a team of $2$ collaborates to compete against the other. In its bidding phase, players bid in turn to find a good contract through a limited information channel. Based on a strong baseline agent that bids competitive bridge purely through domain-agnostic self-play, JPS improves collaboration of team players and outperforms WBridge5, a championship-winning software, by $+0.63$ IMPs (International Matching Points) per board over $1000$ games, substantially better than previous SoTA ($+0.41$ IMPs/b against WBridge5). Note that $+0.1$ IMPs/b is regarded as a nontrivial improvement in Computer Bridge.

High dimensional black-box optimization has broad applications but remains a challenging problem to solve. Given a set of samples xi, yi, building a global model (like Bayesian Optimization (BO)) suffers from the curse of dimensionality in the high-dimensional search space, while a greedy search may lead to sub-optimality. By recursively splitting the search space into regions with high/low function values, recent works like LaNAS shows good performance in Neural Architecture Search (NAS), reducing the sample complexity empirically. In this paper, we coin LA-MCTS that extends LaNAS to other domains. Unlike previous approaches, LA-MCTS learns the partition of the search space using a few samples and their function values in an online fashion. While LaNAS uses linear partition and performs uniform sampling in each region, our LA-MCTS adopts a nonlinear decision boundary and learns a local model to pick good candidates. If the nonlinear partition function and the local model fits well with ground-truth black-box function, then good partitions and candidates can be reached with much fewer samples. LA-MCTS serves as a meta-algorithm by using existing black-box optimizers (e.g., BO, TuRBO as its local models, achieving strong performance in general black-box optimization and reinforcement learning benchmarks, in particular for high-dimensional problems.

Jin-Hwa Kim, Nikita Kitaev, Xinlei Chen, Marcus Rohrbach, Byoung-Tak Zhang, Yuandong Tian, Dhruv Batra, Devi Parikh. Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics. 2019.

Designing accurate and efficient ConvNets for mobile devices is challenging because the design space is combinatorially large. Due to this, previous neural architecture search (NAS) methods are computationally expensive. ConvNet architecture optimality depends on factors such as input resolution and target devices. However, existing approaches are too resource demanding for case-by-case redesigns. Also, previous work focuses primarily on reducing FLOPs, but FLOP count does not always reflect actual latency. To address these, we propose a differentiable neural architecture search (DNAS) framework that uses gradient-based methods to optimize ConvNet architectures, avoiding enumerating and training individual architectures separately as in previous methods. FBNets (Facebook-Berkeley-Nets), a family of models discovered by DNAS surpass state-of-the-art models both designed manually and generated automatically. FBNet-B achieves 74.1% top-1 accuracy on ImageNet with 295M FLOPs and 23.1 ms latency on a Samsung S8 phone, 2.4x smaller and 1.5x faster than MobileNetV2-1.3 with similar accuracy. Despite higher accuracy and lower latency than MnasNet, we estimate FBNet-B's search cost is 420x smaller than MnasNet's, at only 216 GPU-hours. Searched for different resolutions and channel sizes, FBNets achieve 1.5% to 6.4% higher accuracy than MobileNetV2. The smallest FBNet achieves 50.2% accuracy and 2.9 ms latency (345 frames per second) on a Samsung S8. Over a Samsung-optimized FBNet, the iPhone-X-optimized model achieves a 1.4x speedup on an iPhone X. FBNet models are open-sourced at https://github. com/facebookresearch/mobile-vision.

We introduce a new memory architecture, Bayesian Relational Memory (BRM), to improve the generalization ability for semantic visual navigation agents in unseen environments, where an agent is given a semantic target to navigate towards. BRM takes the form of a probabilistic relation graph over semantic entities (e.g., room types), which allows (1) capturing the layout prior from training environments, i.e., prior knowledge, (2) estimating posterior layout at test time, i.e., memory update, and (3) efficient planning for navigation, altogether. We develop a BRM agent consisting of a BRM module for producing sub-goals and a goal-conditioned locomotion module for control. When testing in unseen environments, the BRM agent outperforms baselines that do not explicitly utilize the probabilistic relational memory structure.

The AlphaGo, AlphaGo Zero, and AlphaZero series of algorithms are remarkable demonstrations of deep reinforcement learning’s capabilities, achieving superhuman performance in the complex game of Go with progressively increasing autonomy. However, many obstacles remain in the understanding of and usability of these promising approaches by the research community. Toward elucidating unresolved mysteries and facilitating future research, we propose ELF OpenGo, an open-source reimplementation of the AlphaZero algorithm. ELF OpenGo is the first open-source Go AI to convincingly demonstrate superhuman performance with a perfect (20:0) record against global top professionals. We apply ELF OpenGo to conduct extensive ablation studies, and to identify and analyze numerous interesting phenomena in both the model training and in the gameplay inference procedures. Our code, models, selfplay datasets, and auxiliary data are publicly available.

Search-based methods for hard combinatorial optimization are often guided by heuristics. Tuning heuristics in various conditions and situations is often time-consuming. In this paper, we propose NeuRewriter that learns a policy to pick heuristics and rewrite the local components of the current solution to iteratively improve it until convergence. The policy factorizes into a region-picking and a rule-picking component, each parameterized by a neural network trained with actor-critic methods in reinforcement learning. NeuRewriter captures the general structure of combinatorial problems and shows strong performance in three versatile tasks: expression simplification, online job scheduling and vehicle routing problems. NeuRewriter outperforms the expression simplification component in Z3; outperforms DeepRM and Google OR-tools in online job scheduling; and outperforms recent neural baselines and Google OR-tools in vehicle routing problems.

Reverse engineering of binary executables is a critical problem in the computer security domain. On the one hand, malicious parties may recover interpretable source codes from the software products to gain commercial advantages. On the other hand, binary decompilation can be leveraged for code vulnerability analysis and malware detection. However, efficient binary decompilation is challenging. Conventional decompilers have the following major limitations: (i) they are only applicable to specific source-target language pair, hence incurs undesired development cost for new language tasks; (ii) their output high-level code cannot effectively preserve the correct functionality of the input binary; (iii) their output program does not capture the semantics of the input and the reversed program is hard to interpret. To address the above problems, we propose Coda1, the first end-to-end neural-based framework for code decompilation. Coda decomposes the decompilation task into of two key phases: First, Coda employs an instruction type-aware encoder and a tree decoder for generating an abstract syntax tree (AST) with attention feeding during the code sketch generation stage. Second, Coda then updates the code sketch using an iterative error correction machine guided by an ensembled neural error predictor. By finding a good approximate candidate and then fixing it towards perfect, Coda achieves superior with performance compared to baseline approaches. We assess Coda’s performance with extensive experiments on various benchmarks. Evaluation results show that Coda achieves an average of 82% program recovery accuracy on unseen binary samples, where the state-of-the-art decompilers yield 0% accuracy. Furthermore, Coda outperforms the sequence-to-sequence model with attention by a margin of 70% program accuracy. Our work reveals the vulnerability of binary executables and imposes a new threat to the protection of Intellectual Property (IP) for software development.

We explore using latent natural language instructions as an expressive and compositional representation of complex actions for hierarchical decision making. Rather than directly selecting micro-actions, our agent first generates a latent plan in natural language, which is then executed by a separate model. We introduce a challenging real-time strategy game environment in which the actions of a large number of units must be coordinated across long time scales. We gather a dataset of 76 thousand pairs of instructions and executions from human play, and train instructor and executor models. Experiments show that models using natural language as a latent variable significantly outperform models that directly imitate human actions. The compositional structure of language proves crucial to its effectiveness for action representation. We also release our code, models and data.

The success of lottery ticket initializations (Frankle and Carbin, 2019) suggests that small, sparsified networks can be trained so long as the network is initialized appropriately. Unfortunately, finding these "winning ticket'' initializations is computationally expensive. One potential solution is to reuse the same winning tickets across a variety of datasets and optimizers. However, the generality of winning ticket initializations remains unclear. Here, we attempt to answer this question by generating winning tickets for one training configuration (optimizer and dataset) and evaluating their performance on another configuration. Perhaps surprisingly, we found that, within the natural images domain, winning ticket initializations generalized across a variety of datasets, including Fashion MNIST, SVHN, CIFAR-10/100, ImageNet, and Places365, often achieving performance close to that of winning tickets generated on the same dataset. Moreover, winning tickets generated using larger datasets consistently transferred better than those generated using smaller datasets. We also found that winning ticket initializations generalize across optimizers with high performance. These results suggest that winning ticket initializations generated by sufficiently large datasets contain inductive biases generic to neural networks more broadly which improve training across many settings and provide hope for the development of better initialization methods.

We consider the problem of learning an one-hidden-layer neural network with non-overlapping convolutional layer and ReLU activation function, i.e., $f(Z; w, a) = \sum_j a_j\sigma(w^\top Z_j)$, in which both the convolutional weights $w$ and the output weights $a$ are parameters to be learned. We prove that with Gaussian input $\mathbf{Z}$ there is a spurious local minimizer. Surprisingly, in the presence of the spurious local minimizer, starting from randomly initialized weights, gradient descent with weight normalization can still be proven to recover the true parameters with constant probability (which can be boosted to probability $1$ with multiple restarts). We also show that with constant probability, the same procedure could also converge to the spurious local minimum, showing that the local minimum plays a non-trivial role in the dynamics of gradient descent. Furthermore, a quantitative analysis shows that the gradient descent dynamics has two phases: it starts off slow, but converges much faster after several iterations.

Despite recent successes in synthesizing faces and bedrooms, existing generative models struggle to capture more complex image types (Figure 1), potentially due to the oversimplification of their latent space constructions. To tackle this issue, building on Variational Autoencoders (VAEs), we integrate recurrent connections across channels to both inference and generation steps, allowing the high-level features to be captured in global-to-local, coarse-to-fine manners. Combined with adversarial loss, our channel-recurrent VAE-GAN (crVAE-GAN) outperforms VAE-GAN in generating a diverse spectrum of high resolution images while maintaining the same level of computational efficacy. Our model produces interpretable and expressive latent representations to benefit downstream tasks such as image completion. Moreover, we propose two novel regularizations, namely the KL objective weighting scheme over time steps and mutual information maximization between transformed latent variables and the outputs, to enhance the training.

Deep learning (DL) and reinforcement learning (RL) have been applied with great success to many games, including Go and Atari 2600 games. Monte Carlo Tree Search (MCTS), developed in 2006, can be viewed as a kind of online RL. This technique has greatly improved the level of Go-playing programs. MCTS has since become the state of the art for many other games including Hex, Havannah, and general game playing, and has found much success in applications as diverse as scheduling, unit commitment problems, and probabilistic planning. DL has transformed fields such as image and video recognition and speech understanding. In computer games, DL started making its mark in 2014, when teams from the University of Edinburgh and Google DeepMind independently applied deep convolutional neural networks (DCNNs) to the problem of expertmove prediction in Go.Clark and Storkey’s DCNN achieved a move prediction rate of 44%, exceeding all previously published results. DeepMind’s publication followed soon after, with a DCNN that reached 55%. The combination of DL and RL led to great advances in Atari 2600 game playing, and to the ultimate breakthrough in computer Go. In 2017, DeepMind proposed a new deep reinforcement learning (DRL) algorithm and developed AlphaGo Zero, which is significant for not requiring any human knowledge of Go. By removing the requirement for domain knowledge, DRL is also flexible in that the method can be applied to a wide range of games and problems, ushering in a variety of new research opportunities. In this special issue, we are delighted to bring you eight articles on applying DL/RL related techniques to games research.

Common visual recognition tasks such as classification, object detection, and semantic segmentation are rapidly reaching maturity, and given the recent rate of progress, it is not unreasonable to conjecture that techniques for many of these problems will approach human levels of performance in the next few years. In this paper we look to the future: what is the next frontier in visual recognition? We offer one possible answer to this question. We propose a detailed image annotation that captures information beyond the visible pixels and requires complex reasoning about full scene structure. Specifically, we create an amodal segmentation of each image: the full extent of each region is marked, not just the visible pixels. Annotators outline and name all salient regions in the image and specify a partial depth order. The result is a rich scene structure, including visible and occluded portions of each region, figure-ground edge information, semantic labels, and object overlap. We create two datasets for semantic amodal segmentation. First, we label 500 images in the BSDS dataset with multiple annotators per image, allowing us to study the statistics of human annotations. We show that the proposed full scene annotation is surprisingly consistent between annotators, including for regions and edges. Second, we annotate 5000 images from COCO. This larger dataset allows us to explore a number of algorithmic ideas for amodal segmentation and depth ordering. We introduce novel metrics for these tasks, and along with our strong baselines, define concrete new challenges for the community.

In this paper, we explore theoretical properties of training a two-layered ReLU network $g(\mathbf{x}; \mathbf{w}) = \sum_{j=1}^K \sigma(\mathbf{w}_j^\top\mathbf{x})$ with centered $d$-dimensional spherical Gaussian input $\mathbf{x}$ ($\sigma$=ReLU). We train our network with gradient descent on $\mathbf{w}$ to mimic the output of a teacher network with the same architecture and fixed parameters $\mathbf{w}^*$. We show that its population gradient has an analytical formula, leading to interesting theoretical analysis of critical points and convergence behaviors. First, we prove that critical points outside the hyperplane spanned by the teacher parameters (“out-of-plane“) are not isolated and form manifolds, and characterize in-plane critical-point-free regions for two-ReLU case. On the other hand, convergence to $\mathbf{w}^*$ for one ReLU node is guaranteed with at least $(1-\epsilon)/2$ probability, if weights are initialized randomly with standard deviation upper-bounded by $O(\epsilon/\sqrt{d})$, in accordance with empirical practice. For network with many ReLU nodes, we prove that an infinitesimal perturbation of weight initialization results in convergence towards $\mathbf{w}^*$ (or its permutation), a phenomenon known as spontaneous symmetric-breaking (SSB) in physics. We assume no independence of ReLU activations. Simulation verifies our findings.

In this paper, we propose ELF, an Extensive, Lightweight and Flexible platform for fundamental reinforcement learning research. Using ELF, we implement a highly customizable real-time strategy (RTS) engine with three game environments (Mini-RTS, Capture the Flag and Tower Defense). Mini-RTS, as a miniature version of StarCraft, captures key game dynamics and runs at 165K frame-per-second (FPS) on a laptop. When coupled with modern reinforcement learning methods, the system can train a full-game bot against built-in AIs end-to-end in one day with 6 CPUs and 1 GPU. In addition, our platform is flexible in terms of environment-agent communication topologies, choices of RL methods, changes in game parameters, and can host existing C/C++-based game environments like ALE. Using ELF, we thoroughly explore training parameters and show that a network with Leaky ReLU and Batch Normalization coupled with long-horizon training and progressive curriculum beats the rule-based built-in AI more than 70% of the time in the full game of Mini-RTS. Strong performance is also achieved on the other two games. In game replays, we show our agents learn interesting strategies. ELF, along with its RL platform, is open-sourced at https://github.com/facebookresearch/ELF.

Real-world surfaces such as clothing, water and human body deform in complex ways. The image distortions observed are high-dimensional and non-linear, making it hard to estimate these deformations accurately. The recent data-driven descent approach applies Nearest Neighbor estimators iteratively on a particular distribution of training samples to obtain a globally optimal and dense deformation field between a template and a distorted image. In this work, we develop a hierarchical structure for the Nearest Neighbor estimators, each of which can have only a local image support. We demonstrate in both theory and practice that this algorithm has several advantages over the non-hierarchical version: it guarantees global optimality with significantly fewer training samples, is several orders faster, provides a metric to decide whether a given image is ``hard'' (or ``easy'') requiring more (or less) samples, and can handle more complex scenes that include both global motion and local deformation. The proposed algorithm successfully tracks a broad range of non-rigid scenes including water, clothing, and medical images, and compares favorably against several other deformation estimation and tracking approaches that do not provide optimality guarantees.

Turbulence near hot surfaces such as desert terrains and roads during the summer, causes shimmering, distortion and blurring in images. While recent works have focused on image restoration, this paper explores what information about the scene can be extracted from the distortion caused by turbulence. Based on the physical model of wave propagation, we first study the relationship between the scene depth and the amount of distortion caused by homogenous turbulence. We then extend this relationship to more practical scenarios such as finite extent and height-varying turbulence, and present simple algorithms to estimate depth ordering, depth discontinuity and relative depth, from a sequence of short exposure images. In the case of general non-homogenous turbulence, we show that a statistical property of turbulence can be used to improve long-range structure-from-motion (or stereo). We demonstrate the accuracy of our methods in both laboratory and outdoor settings and conclude that turbulence (when present) can be a strong and useful depth cue.

Crowdsourcing has recently become popular among machine learning researchers and social scientists as an effective way to collect large-scale experimental data from distributed workers. To extract useful information from the cheap but potentially unreliable answers to tasks, a key problem is to identify reliable workers as well as unambiguous tasks. Although for objective tasks that have one correct answer per task, previous works can estimate worker reliability and task clarity based on the single gold standard assumption, for tasks that are subjective and accept multiple reasonable answers that workers may be grouped into, a phenomenon called schools of thought, existing models cannot be trivially applied. In this work, we present a statistical model to estimate worker reliability and task clarity without resorting to the single gold standard assumption. This is instantiated by explicitly characterizing the grouping behavior to form schools of thought with a rank-1 factorization of a worker-task groupsize matrix. Instead of performing an intermediate inference step, which can be expensive and unstable, we present an algorithm to analytically compute the sizes of different groups. We perform extensive empirical studies on real data collected from Amazon Mechanical Turk. Our method discovers the schools of thought, shows reasonable estimation of worker reliability and task clarity, and is robust to hyperparameter changes. Furthermore, our estimated worker reliability can be used to improve the gold standard prediction for objective tasks.

Kernel methods have been popular over the last decade to solve many computer vision, statistics and machine learning problems. An important, both theoretically and practically, open problem in kernel methods is the pre-image problem. The pre-image problem consists of finding a vector in the input space whose mapping is known in the feature space induced by a kernel. To solve the pre-image problem, this paper proposes a framework that computes an isomorphism between local Gram matrices in the input and feature space. Unlike existing methods that rely on analytic properties of kernels, our framework derives closed-form solutions to the pre-image problem in the case of non-differentiable and application-specific kernels. Experiments on the pre-image problem for visualizing cluster centers computed by kernel k-means and denoising high-dimensional images show that our algorithm outperforms state-of-the-art methods.

Distortions in images of documents, such as the pages of books, adversely affect the performance of optical character recognition (OCR) systems. Removing such distortions requires the 3D deformation of the document that is often measured using special and precisely calibrated hardware (stereo, laser range scanning or structured light). In this paper, we introduce a new approach that automatically reconstructs the 3D shape and rectifies a deformed text document from a single image. We first estimate the 2D distortion grid in an image by exploiting the line structure and stroke statistics in text documents. This approach does not rely on more noise-sensitive operations such as image binarization and character segmentation. The regularity in the text pattern is used to constrain the 2D distortion grid to be a perspective projection of a 3D parallelogram mesh. Based on this constraint, we present a new shape-from-texture method that computes the 3D deformation up to a scale factor using SVD. Unlike previous work, this formulation imposes no restrictions on the shape (e.g., a developable surface). The estimated shape is then used to remove both geometric distortions and photometric (shading) effects in the image. We demonstrate our techniques on documents containing a variety of languages, fonts and sizes.

Image alignment in the presence of non-rigid distortions is a challenging task. Typically, this involves estimating the parameters of a dense deformation field that warps a distorted image back to its undistorted template. Generative approaches based on parameter optimization such as Lucas-Kanade can get trapped within local minima. On the other hand, discriminative approaches like Nearest-Neighbor require a large number of training samples that grows exponentially with the desired accuracy. In this work, we develop a novel data-driven iterative algorithm that combines the best of both generative and discriminative approaches. For this, we introduce the notion of a “pull-back” operation that enables us to predict the parameters of the test image using training samples that are not in its neighborhood (not ϵ-close) in parameter space. We prove that our algorithm converges to the global optimum using a significantly lower number of training samples that grows only logarithmically with the desired accuracy. We analyze the behavior of our algorithm extensively using synthetic data and demonstrate successful results on experiments with complex deformations due to water and clothing.

Most active scene recovery techniques assume that a scene point is illuminated only directly by the illumination source. Consequently, global illumination effects due to inter-reflections, sub-surface scattering and volumetric scattering introduce strong biases in the recovered scene shape. Our goal is to recover scene properties in the presence of global illumination. To this end, we study the interplay between global illumination and the depth cue of illumination defocus. By expressing both these effects as low pass filters, we derive an approximate invariant that can be used to separate them without explicitly modeling the light transport. This is directly useful in any scenario where limited depth-of-field devices (such as projectors) are used to illuminate scenes with global light transport and significant depth variations. We show two applications: (a) accurate depth recovery in the presence of global illumination, and (b) factoring out the effects of defocus for correct direct-global separation in large depth scenes. We demonstrate our approach using scenes with complex shapes, reflectances, textures and translucencies.

A video sequence of an underwater scene taken from above the water surface suffers from severe distortions due to water fluctuations. In this paper, we simultaneously estimate the shape of the water surface and recover the planar underwater scene without using any calibration patterns, image priors, multiple viewpoints or active illumination. The key idea is to build a compact spatial distortion model of the water surface using the wave equation. Based on this model, we present a novel tracking technique that is designed specifically for water surfaces and addresses two unique challenges—the absence of an object model or template and the presence of complex appearance changes in the scene due to water fluctuation. We show the effectiveness of our approach on both simulated and real scenes, with text and texture.

A family photo album based cartoon personalization system, EasyToon, is proposed in this paper. Using state of the art computer vision and graphics technologies and effective UI design, the interactive tool can quickly generate a personalized cartoon storyboard, which naturally blends a real face chosen from the family photo album into a cartoon picture. The personalized cartoon image is easily and quickly obtained in two main steps. First, the best face candidate is selected from the album interactively. Then a personalized cartoon image is automatically synthesized by blending the selected face into the interesting cartoon image. Experiments show that most users express great interest in our system. Without any art background, they can make a personalized cartoon of high quality using the EasyToon within minutes.

Digital photo management is becoming indispensable for the explosively growing family photo albums due to the rapid popularization of digital cameras and mobile phone cameras. In an effective photo management system photo annotation is the most challenging task. In this paper, we develop several innovative interaction techniques for semi-automatic photo annotation. Compared with traditional annotation systems, our approach provides the following new features: "cluster annotation" puts similar faces or photos with similar scene together, and enables user label them in one operation; "contextual re-ranking" boosts the labeling productivity by guessing the user intention; "ad hoc annotation" allows user label photos while they are browsing or searching, and improves system performance progressively through learning propagation. Our results show that these technologies provide a more user friendly interface for the annotation of person name, location, and event, and thus substantially improve the annotation performance especially for a large photo album.

Face annotation technology is important for a photo management system. In this paper, we propose a novel interactive face annotation framework combining unsupervised and interactive learning. There are two main contributions in our framework. In the unsupervised stage, a partial clustering algorithm is proposed to find the most evident clusters instead of grouping all instances into clusters, which leads to a good initial labeling for later user interaction. In the interactive stage, an efficient labeling procedure based on minimization of both global system uncertainty and estimated number of user operations is proposed to reduce user interaction as much as possible. Experimental results show that the proposed annotation framework can significantly reduce the face annotation workload and is superior to existing solutions in the literature.

A fundamental challenge in face recognition lies in determining what facial features are important for the identification of faces. In this paper, a novel face recognition framework is proposed to address this problem. In our framework, 3D face models are used to synthesize a huge database of realistic face images which covers wide appearance variations of faces due to various pose, illumination, and expression changes. A novel feature selection algorithm which we call Joint Boosting is developed to extract discriminative face features using this massive database. The major contributions of this paper are: (1) With the help of 3D face models, a massive database of realistic virtual face images is generated to achieve robust feature selection; (2)Because the huge database covers a wide range of face variations, our feature selection procedure only needs to be trained once, and the selected feature set can be generalized to other face database without re-training; (3) We propose a new learning algorithm, Joint Boosting Algorithm, which is effective and efficient in learning directly from a massive database without having to convert face images to intra-personal and extra-personal difference images. This property is important for applying our algorithm to other general pattern recognition problems. Experimental results show that our method significantly improves recognition performance.