Decoder-based large language models (LLMs) have proven highly versatile, with remarkable successes even on problems ostensibly removed from traditional language generation. One such example is solving regression problems, where the targets are real numbers rather than textual tokens. A common approach to use LLMs on such problems is to perform fine-tuning based on the cross-entropy loss, and use autoregressive sampling at inference time. Another approach relies on fine-tuning a separate predictive head with a suitable loss such as squared error. While each approach has had success, there has been limited study on principled ways of using decoder LLMs for regression. In this work, we compare different prior works under a unified view, and introduce regression-aware fine-tuning(RAFT), a novel approach based on the Bayes-optimal decision rule. We demonstrate how RAFT improves over established baselines on several benchmarks and model families.

Cascades and speculative decoding are two common approaches to improving language models' inference efficiency. Both approaches interleave two models, but via fundamentally distinct mechanisms: deferral rule that invokes the larger model only for “hard” inputs, while speculative decoding uses speculative execution to primarily invoke the larger model in parallel scoring mode. These mechanisms offer different benefits: empirically, cascades offer compelling cost-quality trade-offs, often even outperforming the large model; speculative cascades offer impressive speed-ups, while guaranteeing quality-neutrality. In this paper, we leverage the best of both these approaches by designing new speculative cascading techniques that implement their deferral rule through speculative execution. We characterize the optimal deferral rule for our speculative cascades, and employ a plug-in approximation to the optimal rule. Experiments with Gemma and T5 models on a range of language benchmarks show that our approach yields better cost quality trade-offs than cascading and speculative decoding baselines.

Recent developments in large language models have sparked interest in efficient pretraining methods. Stagewise training approaches to improve efficiency, like gradual stacking and layer dropping (Reddi et al., 2023; Zhang & He, 2020), have recently garnered attention. The prevailing view suggests that stagewise dropping strategies, such as layer dropping, are ineffective, especially when compared to stacking-based approaches. This paper challenges this notion by demonstrating that, with proper design, dropping strategies can be competitive, if not better, than stacking methods. Specifically, we develop a principled stagewise training framework, progressive subnetwork training, which only trains subnetworks within the model and progressively increases the size of subnetworks during training, until it trains the full network. We propose an instantiation of this framework — Random Part Training (RAPTR) — that selects and trains only a random subnetwork (e.g. depth-wise, width-wise) of the network at each step, progressively increasing the size in stages. We show that this approach not only generalizes prior works like layer dropping but also fixes their key issues. Furthermore, we establish a theoretical basis for such approaches and provide justification for (a) increasing complexity of subnetworks in stages, conceptually diverging from prior works on layer dropping, and (b) stability in loss across stage transitions in presence of key modern architecture components like residual connections and layer norms. Through comprehensive experiments, we demonstrate that RAPTR can significantly speed up training of standard benchmarks like BERT and UL2, up to 33% compared to standard training and, surprisingly, also shows better downstream performance on UL2, improving QA tasks and SuperGLUE by 1.5%; thereby, providing evidence of better inductive bias.

Large language models have shown remarkable reasoning abilities and scaling laws suggest that large parameter count, especially along the depth axis, is the primary driver. In this work, we make a stronger claim --- many reasoning problems require a large depth but not necessarily many parameters. This unlocks a novel application of looped models for reasoning. Firstly, we show that for many synthetic reasoning problems like addition, $p$-hop induction, and math problems, a $k$-layer transformer looped $L$ times nearly matches the performance of a $kL$-layer non-looped model, and is significantly better than a $k$-layer model. This is further corroborated by theoretical results showing that many such reasoning problems can be solved via iterative algorithms, and thus, can be solved effectively using looped models with nearly optimal depth. Perhaps surprisingly, these benefits also translate to practical settings of language modeling --- on many downstream reasoning tasks, a language model with $k$-layers looped $L$ times can be competitive to, if not better than, a $kL$-layer language model. In fact, our empirical analysis reveals an intriguing phenomenon: looped and non-looped models exhibit scaling behavior that depends on their effective depth, akin to the inference-time scaling of chain-of-thought (CoT) reasoning. We further elucidate the connection to CoT reasoning by proving that looped models implicitly generate latent thoughts and can simulate $T$ steps of CoT with $T$ loops. Inspired by these findings, we also present an interesting dichotomy between reasoning and memorization, and design a looping-based regularization that is effective on both fronts.

Low-rank adaption (LoRA) is a widely used parameter-efficient finetuning method for LLM that reduces memory requirements. However, current LoRA optimizers lack transformation invariance, meaning the updates depending on how the two LoRA factors are scaled or rotated. This deficiency leads to inefficient learning and sub-optimal solutions in practice. This paper introduces LoRA-RITE, a novel adaptive matrix preconditioning method for LoRA optimization, which can achieve transformation invariance and remain computationally efficient. We provide theoretical analysis to demonstrate the benefit of our method and conduct experiments on various LLM tasks with different models including Gemma 2B, 7B, and mT5-XXL. The results demonstrate consistent improvements against existing optimizers. For example, replacing Adam with LoRA-RITE during LoRA fine-tuning of Gemma-2B yielded 4.6% accuracy gain on Super-Natural Instructions and 3.5% accuracy gain across other four LLM benchmarks (HellaSwag, ArcChallenge, GSM8K, OpenBookQA).

Bipartite ranking is a fundamental supervised learning problem, with the goal of learning a ranking over instances with maximal area under the ROC curve (AUC) against a single binary target label. However, one may often observe multiple binary target labels, e.g., from distinct human annotators. How can one synthesize such labels into a single coherent ranking? In this work, we formally analyze two approaches to this problem—loss aggregation and label aggregation—by characterizing their Bayes-optimal solutions. We show that while both approaches can yield Pareto-optimal solutions, loss aggregation can exhibit label dictatorship: one can inadvertently (and undesirably) favor one label over others. This suggests that label aggregation can be preferable to loss aggregation, which we empirically verify.

One of the core pillars of efficient deep learning methods are architectural improvements, such as residual/skip connections, which have led to significantly better model convergence and quality. Since their introduction, residual connections have become ubiquitous not only in convolutional neural networks but also in transformer-based architectures, the backbone of LLMs. In this paper, we introduce the Learned Augmented Residual Layer (LAuReL) — a novel generalization of the canonical residual connection — designed to serve as an in-situ replacement while outperforming it in both model quality and footprint metrics. Our experiments show that LAuReL can enhance quality for both vision and language models while adding fewer parameters and incurring less latency and memory overhead than naively increasing parameter count. For example, on the ImageNet-1K task, LAuReL achieves the same model quality improvements as naively adding an extra layer while using $2.6 \times$ fewer parameters. Similarly, when pre-training 1B and 4B parameter LLMs, LAuReL improves performance on a variety of challenging downstream evaluation tasks by 2.54% to 20.05%, while adding only 0.012% and 0.1% additional parameters, respectively.

We present a novel unified analysis for a broad class of adaptive optimization algorithms with structured (e.g., layerwise, diagonal, and kronecker-factored) preconditioners for both online regret minimization and offline convex optimization. Our analysis not only provides matching rate to several important structured preconditioned algorithms including diagonal AdaGrad, full-matrix AdaGrad, and AdaGrad-Norm, but also gives an improved convergence rate for a one-sided variant of Shampoo over that of original Shampoo. Interestingly, more structured preconditioners (e.g., diagonal Adagrad, AdaGrad-Norm which use less space and compute) are often presented as computationally efficient approximations to full-matrix Adagrad, aiming for improved optimization performance through better approximations. Our unified analysis challenges this prevailing view and reveals, perhaps surprisingly, that more structured preconditioners, despite using less space and computation per step, can outperform their less structured counterparts. To demonstrate this, we show that one-sided Shampoo, which is relatively much cheaper than full-matrix AdaGrad could outperform it both theoretically and experimentally.

In-context Ranking (ICR) is an emerging paradigm for Information Retrieval (IR), which leverages contextual understanding of LLMs by directly incorporating the task description, candidate documents, and the query into the model's input prompt and tasking the LLM to identify relevant document(s). While it is effective, efficiency is a significant challenge in this paradigm, especially as the candidate list grows due to quadratic/super-linear scaling of attention operation with context length. To this end, this paper first identifies inherent and exploitable structures in the attention of LLMs finetuned for ICR: (1) inter-document block sparsity: attention is dense within each document block but sparse across different documents in the context; and (2) query-document block relevance: the attention scores from certain query tokens to a document block in middle layers strongly correlate with that document's actual relevance. Motivated by these observations, we introduce BlockRank (Blockwise In-context Ranking), a novel method that adapts the attention operation in an LLM by (a) architecturally enforcing the observed inter-document block sparsity, reducing attention complexity from quadratic to linear without loss in performance, and (b) optimizing query-document block relevance for true relevant documents during fine-tuning using an auxiliary contrastive training objective, improving retrieval in attention. Experiments on BEIR, MSMarco and NQ with Mistral-7B demonstrate that BlockRank Mistral matches or outperforms existing SOTA listwise rankers and controlled fine-tuned baseline while being significantly more efficient at inference (4.7x for 100 MSMarco documents in context) and scaling gracefully to long-context shortlists, around 500 documents in-context (approximately 100K context length) within a second, presenting a scalable and effective solution for ICR.

Measuring similarity between training examples is critical for curating high-quality and diverse pretraining datasets for language models. However, similarity is typically computed with a generic off-the-shelf embedding model that has been trained for tasks such as retrieval. Whether these embedding-based similarity metrics are well-suited for pretraining data selection remains largely unexplored. In this paper, we propose a new framework to assess the suitability of a similarity metric specifically for data curation in language model pretraining applications. Our framework's first evaluation criterion captures how well distances reflect generalization in pretraining loss between different training examples. Next, we use each embedding model to guide a standard diversity-based data curation algorithm and measure its utility by pretraining a language model on the selected data and evaluating downstream task performance. Finally, we evaluate the capabilities of embeddings to distinguish between examples from different data sources. With these evaluations, we demonstrate that standard off-the-shelf embedding models are not well-suited for the pretraining data curation setting, underperforming even remarkably simple embeddings that are extracted from models trained on the same pretraining corpus. Our experiments are performed on the Pile, for pretraining a 1.7B parameter language model on 200B tokens. We believe our analysis and evaluation framework serves as a foundation for the future design of embeddings that specifically reason about similarity in pretraining datasets.

Dual encoder (DE) models, where a pair of matching query and document are embedded into similar vector representations, are widely used in information retrieval due to their simplicity and scalability. However, the Euclidean geometry of the embedding space limits the expressive power of DEs, which may compromise their quality. This paper investigates such limitations in the context of hierarchical retrieval (HR), where the document set has a hierarchical structure and the matching documents for a query are all of its ancestors. We first prove that DEs are feasible for HR as long as the embedding dimension is linear in the depth of the hierarchy and logarithmic in the number of documents. Then we study the problem of learning such embeddings in a standard retrieval setup where DEs are trained on samples of matching query and document pairs. Our experiments reveal a lost-in-the-long-distance phenomenon, where retrieval accuracy degrades for documents further away in the hierarchy. To address this, we introduce a pretrain-finetune recipe that significantly improves long-distance retrieval without sacrificing performance on closer documents. We experiment on a realistic hierarchy from WordNet for retrieving documents at various levels of abstraction, and show that pretrain-finetune boosts the recall on long-distance pairs from 19% to 76%. Finally, we demonstrate that our method improves retrieval of relevant products on a shopping queries dataset.

The discovery of the *lazy neuron phenomenon* (Li et al., 2022), where fewer than 10% of the feedforward networks (FFN) parameters in trained Transformers are activated per token, has spurred significant interests in *activation sparsity* for enhancing large model efficiency. While notable progress has been made in translating such sparsity to wall-time benefits across CPUs, GPUs, and TPUs, modern Transformers have moved away from the ReLU activation function crucial to this phenomenon. Existing efforts on re-introducing activation sparsity, e.g., by reverting to ReLU or applying top-k masking, often degrade model quality, increase parameter count, or complicate training. Sparse attention, the application of sparse activation to the attention mechanism, often face similar challenges. This paper introduces the Spark Transformer, a novel architecture that achieves high activation sparsity in both FFN and the attention mechanism while maintaining model quality, parameter count, and standard training procedures. Our method realizes sparsity via top-$k$ masking for explicit control over sparsity level. Crucially, we introduce *statistical top-k*, a hardware-accelerator-friendly, linear-time approximate algorithm that avoids costly sorting and mitigates significant training slowdown from standard top-k operators. Furthermore, Spark Transformer reallocates existing FFN parameters and attention key embeddings to form a low-cost predictor for identifying activated entries. This design not only mitigates quality loss from enforced sparsity, but also enhances wall-time benefit. Pretrained with the Gemma-2 recipe, Spark Transformer demonstrates competitive performance on standard benchmarks while exhibiting significant sparsity: only 8\% of FFN neurons are activated, and each token attends to a maximum of 256 tokens. This translates to a 2.5x reduction in FLOPs, leading to decoding wall-time speedups of up to 1.79x on CPU and 1.40xon GPU.

Modern text-to-image generation models produce high-quality images that are both photorealistic and faithful to the text prompts. However, this quality comes at significant computational cost: nearly all of these models are iterative and require running sampling multiple times with large models. This iterative process is needed to ensure that different regions of the image are not only aligned with the text prompt, but also compatible with each other. In this work, we propose a light-weight approach to achieving this compatibility between different regions of an image, using a Markov Random Field (MRF) model. We demonstrate the effectiveness of this method on top of the latent token-based Muse text-to-image model. The MRF richly encodes the compatibility among image tokens at different spatial locations to improve quality and significantly reduce the required number of Muse sampling steps. Inference with the MRF is significantly cheaper, and its parameters can be quickly learned through back-propagation by modeling MRF inference as a differentiable neural-network layer. Our full model, MarkovGen, uses this proposed MRF model to both speed up Muse by 1.5 x and produce higher quality images by decreasing undesirable image artifacts.

As with many machine learning problems, the progress of image generation methods hinges on good evaluation metrics. One of the most popular is the Frechet Inception Distance (FID). FID estimates the distance between a distribution of Inception-v3 features of real images, and those of images generated by the algorithm. We highlight important drawbacks of FID: Inception's poor representation of the rich and varied content generated by modern text-to-image models, incorrect normality assumptions, and poor sample complexity. We call for a reevaluation of FID's use as the primary quality metric for generated images. We empirically demonstrate that FID contradicts human raters, it does not reflect gradual improvement of iterative text-to-image models, it does not capture distortion levels, and that it produces inconsistent results when varying the sample size. We also propose an alternative new metric, CMMD, based on richer CLIP embeddings and the maximum mean discrepancy distance with the Gaussian RBF kernel. It is an unbiased estimator that does not make any assumptions on the probability distribution of the embeddings and is sample efficient. Through extensive experiments and analysis, we demonstrate that FID-based evaluations of text-to-image models may be unreliable, and that CMMD of-fers a more robust and reliable assessment of image quality. A reference implementation of CMMD is available at: https://github.com/google-research/google-research/tree/master/cmmd.

Classical wisdom in machine learning holds that the generalization error can be decomposed into bias and variance, and these two terms exhibit a \emph{trade-off}. However, in this paper, we show that for an ensemble of deep learning based classification models, bias and variance are \emph{aligned} at a sample level, where squared bias is approximately \emph{equal} to variance for correctly classified sample points. We present empirical evidence confirming this phenomenon in a variety of deep learning models and datasets. Moreover, we study this phenomenon from two theoretical perspectives: calibration and neural collapse. We first show theoretically that under the assumption that the models are well calibrated, we can observe the bias-variance alignment. Second, starting from the picture provided by the neural collapse theory, we show an approximate correlation between bias and variance.

Language models generate responses by producing a series of tokens in immediate succession: the $(K+1)^{\rm th}$ token is an outcome of manipulating $K$ hidden vectors per layer, one vector per preceding token. What if instead we were to let the model manipulate say, $K+10$ hidden vectors, before it outputs the $(K+1)^{\rm th}$ token? We operationalize this idea by performing training and inference on language models with a (learnable) $\textit{pause}$ token, a sequence of which is appended to the input prefix. We then delay extracting the model's outputs until the last pause token is seen, thereby allowing the model to process extra computation before committing to an answer. We empirically evaluate $\textit{pause-training}$ on decoder-only models of 1B and 130M parameters with causal pretraining on C4, and on downstream tasks covering reasoning, question-answering, general understanding and fact recall. Our main finding is that inference-time delays show gains when the model is both pre-trained and finetuned with delays. For the 1B model, we witness gains on 8 of 9 tasks, most prominently, a gain of $18\\%$ EM score on the QA task of SQuAD, $8\\%$ on CommonSenseQA and $1\\%$ accuracy on the reasoning task of GSM8k. Our work raises a range of conceptual and practical future research questions on making delayed next-token prediction a widely applicable new paradigm.

Recent advances in language models (LMs) have led to significant improvements in quality on complex NLP tasks, but at the expense of increased inference costs. A simple strategy to achieve more favorable cost-quality tradeoffs is cascading: here, a small model is invoked for most “easy” instances, while a few “hard” instances are deferred to the large model. While the principles underpinning effective cascading are well-studied for classification tasks — with deferral based on predicted class uncertainty favored theoretically and practically — a similar understanding is lacking for generative LM tasks. In this work, we initiate a systematic study of deferral rules for LM cascades. We begin by examining the natural extension of predicted class uncertainty to generative LM tasks, namely, the predicted sequence uncertainty. We show that this measure suffers from the length bias problem, either over- or under-emphasizing outputs based on their lengths. This is because LMs produce a sequence of uncertainty values, one for each output token; and moreover, the number of output tokens is variable across different examples. To mitigate the length bias, we propose to exploit the richer token-level uncertainty information implicit in generative LMs. We argue that naive predicted sequence uncertainty corresponds to a simple aggregation of these uncertainties. By contrast, we show that incorporating token-level uncertainty through learned post-hoc deferral rules can significantly outperform such simple aggregation strategies, via experiments on a range of natural language benchmarks with FLAN-T5 models. We further show that incorporating embeddings from the smaller model and intermediate layers of the larger model can give an additional boost in the overall cost-quality tradeoff.

Preventing the performance decay of Transformers on inputs longer than those used for training has been an important challenge in extending the context length of these models. Though the Transformer architecture has fundamentally no limits on the input sequence lengths it can process, the choice of position encoding used during training can limit the performance of these models on longer inputs. We propose a novel functional relative position encoding with progressive interpolation, FIRE, to improve Transformer generalization to longer contexts. We theoretically prove that this can represent some of the popular relative position encodings, such as T5's RPE, Alibi, and Kerple. We next empirically show that FIRE models have better generalization to longer contexts on both zero-shot language modeling and long text benchmarks.

Learning to reject (L2R) is a classical problem where one seeks a classifier capable of abstaining on low-confidence samples. Most prior work on L2R has focused on minimizing the standard misclassification error. However, in many real-world applications, the label distribution is highly imbalanced, necessitating alternate evaluation metrics such as the balanced error or the worst-group error that enforce equitable performance across both the head and tail classes. In this paper, we establish that traditional L2R methods can be grossly sub-optimal for such metrics, and show that this is due to an intricate dependence in the objective between the label costs and the rejector. We then derive the form of the Bayes-optimal classifier and rejector for the balanced error, propose a novel plug-in approach to mimic this solution, and extend our results to general evaluation metrics. Through experiments on benchmark image classification tasks, we show that our approach yields better trade-offs in both the balanced and worst-group error compared to L2R baselines.

Real-world classifiers can benefit from the option of abstaining from predicting on samples where they have low confidence. Such abstention is particularly useful on samples which are close to the learned decision boundary, or which are outliers with respect to the training sample. These settings have been the subject of extensive but disjoint study in the selective classification (SC) and out-of-distribution (OOD) detection literature. Recent work on selective classification with OOD detection (SCOD) has argued for the unified study of these problems; however, the formal underpinnings of this problem are still nascent, and existing techniques are heuristic in nature. In this paper, we propose new plugin estimators for SCOD that are theoretically grounded, effective, and generalise existing approaches from the SC and OOD detection literature. In the course of our analysis, we formally explicate how naïve use of existing SC and OOD detection baselines may be inadequate for SCOD. We empirically demonstrate that our approaches yields competitive SC and OOD detection trade-offs compared to common baselines.

Pretrained large language models (LLMs) are general purpose problem solvers applicable to a diverse set of tasks with prompts. They can be further improved towards a specific task by fine-tuning on a specialized dataset. However, fine-tuning usually makes the model narrowly specialized on this dataset with reduced general in-context learning performances, which is undesirable whenever the fine-tuned model needs to handle additional tasks where no fine-tuning data is available. In this work, we first demonstrate that fine-tuning on a single task indeed decreases LLMs' general in-context learning performance. We discover one important cause of such forgetting, format specialization, where the model overfits to the format of the fine-tuned task.We further show that format specialization happens at the very beginning of fine-tuning. To solve this problem, we propose Prompt Tuning with MOdel Tuning (ProMoT), a simple yet effective two-stage fine-tuning framework that reduces format specialization and improves generalization.ProMoT offloads task-specific format learning into additional and removable parameters by first doing prompt tuning and then fine-tuning the model itself with this soft prompt attached. With experiments on several fine-tuning tasks and 8 in-context evaluation tasks, we show that ProMoT achieves comparable performance on fine-tuned tasks to standard fine-tuning, but with much less loss of in-context learning performances across a board range of out-of-domain evaluation tasks. More importantly, ProMoT can even enhance generalization on in-context learning tasks that are semantically related to the fine-tuned task, e.g. ProMoT on En-Fr translation significantly improves performance on other language pairs, and ProMoT on NLI improves performance on summarization. Experiments also show that ProMoT can improve the generalization performance of multi-task training.

Speculative decoding~(SD) accelerates large language model inference by employing a faster {\em draft} model for generating multiple tokens, which are then verified in parallel by the larger {\em target} model, resulting in the text generated according to the target model distribution. However, identifying a compact draft model that is well-aligned with the target model is challenging. To tackle this issue, we propose {\em DistillSpec} that uses knowledge distillation to better align the draft model with the target model, before applying SD. DistillSpec makes two key design choices, which we demonstrate via systematic study to be crucial to improve the draft and target alignment: utilizing \emph{on-policy} data generation from the draft model, and \emph{tailoring the divergence function} to the task and decoding strategy. Notably, DistillSpec yields impressive $10 - 45\%$ speedups over standard SD on a range of standard benchmarks, using both greedy and non-greedy sampling. Furthermore, we combine DistillSpec with lossy SD to achieve fine-grained control over the latency vs. task performance trade-off. Finally, in practical scenarios with models of varying sizes, first using distillation to boost the performance of the target model and then applying DistillSpec to train a well-aligned draft model can reduce decoding latency by $6 - 10\times$ with minimal performance drop, compared to standard decoding without distillation.

Transformers to do reasoning and few-shot learning, without any fine-tuning, is widely conjectured to stem from their ability to implicitly simulate a multi-step algorithms – such as gradient descent – with their weights in a single forward pass. Recently, there has been progress in understanding this complex phenomenon from an expressivity point of view, by demonstrating that Transformers can express such multi-step algorithms. However, our knowledge about the more fundamental aspect of its learnability, beyond single layer models, is very limited. In particular, can training Transformers enable convergence to algorithmic solutions? In this work we resolve this for in context linear regression with linear looped Transformers – a multi-layer model with weight sharing that is conjectured to have an inductive bias to learn fix-point iterative algorithms. More specifically, for this setting we show that the global minimizer of the population training loss implements multi-step preconditioned gradient descent, with a preconditioner that adapts to the data distribution. Furthermore, we show a fast convergence for gradient flow on the regression loss, despite the non-convexity of the landscape, by proving a novel gradient dominance condition. To our knowledge, this is the first theoretical analysis for multi-layer Transformer in this setting. We further validate our theoretical findings through synthetic experiments.

Modern information retrieval (IR) systems consists of multiple stages like retrieval and ranking, with Transformer-based models achieving state-of-the-art performance at each stage. In this paper, we challenge the tradition of using separate models for different stages and ask if a single Transformer encoder can provide relevance score needed in each stage. We present USTAD – a new unified approach to train a single network that can provide powerful ranking scores as a cross-encoder (CE) model as well as factorized embeddings for large-scale retrieval as a dual-encoder (DE) model. Empirically, we find a single USTAD model to be competitive to separate ranking CE and retrieval DE models. Furthermore, USTAD combines well with a novel embedding matching-based distillation, significantly improving CE to DE distillation. It further motivates novel asymmetric architectures for student models to ensure a better embedding alignment between the student and the teacher while ensuring small online inference cost. On standard benchmarks like MSMARCO, we demonstrate that USTAD with our proposed distillation method leads to asymmetric students with only 1/10th trainable parameter but retaining 95-97% of the teacher performance.

Autoregressive language models are the currently dominant paradigm for text generation, however they have some fundamental limitations that cannot be remedied by scale—for example inherently sequential and unidirectional generation. While alternate classes of models have been explored, we have limited mathematical understanding of their fundamental power and limitations. In this paper we focus on Generative Masked Language Models (GMLMs), a non-autoregressive paradigm in which we train a model to fit conditional probabilities of the data distribution via masking, which are subsequently used as inputs to a Markov Chain to draw samples from the model. These models empirically strike a promising speed-quality trade-off as each step can be typically parallelized by decoding the entire sequence in parallel. We develop a mathematical framework for analyzing and improving such models which sheds light on questions of sample complexity and inference speed and quality. Empirically, we adapt the T5 model for iteratively-refined parallel decoding, achieving 2-3x speedup in machine translation with minimal sacrifice in quality compared with autoregressive models. We run careful ablation experiments to give recommendations on key design choices, and make fine-grained observations on the common error modes in connection with our theory. Our mathematical analyses and empirical observations characterize both potentials and limitations of this approach, and can be applied to future works on improving understanding and performance of GMLMs. We released codes for our experiments.

The autoregressive nature of conventional large language models (LLMs) inherently limits inference speed, as tokens are generated sequentially. While speculative (Leviathan et al., 2023) and parallel (Stern et al., 2018) decoding techniques attempt to mitigate this, they face limitations: either relying on less accurate smaller models for generation or failing to fully leverage the base LLM’s representations. We introduce a novel architecture, Tandem transformers, to address these issues. This architecture uniquely combines (1) a small autoregressive model and (2) a large model operating in block mode (processing multiple tokens simultaneously). The small model’s predictive accuracy is substantially enhanced by granting it attention to the large model’s richer representations. On the PaLM2 pretraining dataset, a tandem of PaLM2-Bison and PaLM2-Gecko demonstrates a 3.3% improvement in next-token prediction accuracy over a standalone PaLM2-Gecko, offering a 1.16x speedup compared to a PaLM2-Otter model with comparable downstream performance. We further incorporate the Tandem model within the speculative decoding (SPEED) framework where the large model validates tokens from the small model. This ensures that the tandem of PaLM2-Bison and PaLM2-Gecko achieves substantial speedup (around 1.14x faster than using vanilla PaLM2-Gecko in SPEED) while maintaining identical downstream task accuracy.

Autoregressive language models have achieved remarkable advancements, yet their potential is often limited by the slow inference speeds associated with sequential token generation. Blockwise parallel decoding (BPD) was proposed by Stern et al. [42] as a method to improve inference speed of language models by simultaneously predicting multiple future tokens, termed block drafts, which are subsequently verified by the autoregressive model. This paper advances the understanding and improvement of block drafts in two ways. First, we analyze token distributions generated across multiple prediction heads. Second, leveraging these insights, we propose algorithms to improve BPD inference speed by refining the block drafts using task-independent \ngram and neural language models as lightweight rescorers. Experiments demonstrate that by refining block drafts of open-sourced Vicuna and Medusa LLMs, the mean accepted token length are increased by 5-25% relative. This results in over a 3x speedup in wall clock time compared to standard autoregressive decoding in open-source 7B and 13B LLMs.

Given the increasing scale of model sizes, efficient training strategies like gradual stacking have garnered interest. Stacking enables efficient training by gradually growing the depth of a model in stages and using layers from a smaller model in an earlier stage to initialize the next stage. Although efficient for training, the model biases induced by such growing approaches are largely unexplored. In this work, we examine this fundamental aspect of gradual stacking, going beyond its efficiency benefits. We propose a variant of gradual stacking called MIDAS that can speed up language model training by up to 40\%. Furthermore we discover an intriguing phenomenon: MIDAS is not only training-efficient but surprisingly also has an inductive bias towards improving downstream tasks, especially tasks that require reasoning abilities like reading comprehension and math problems, despite having similar or slightly worse perplexity compared to baseline training. To further analyze this inductive bias, we construct {\em reasoning primitives} – simple synthetic tasks that are building blocks for reasoning – and find that a model pretrained with stacking is significantly better than standard pretraining on these primitives, with and without fine-tuning. This provides stronger and more robust evidence for this inductive bias towards reasoning. These findings of training efficiency and inductive bias towards reasoning are verified at 1B, 2B and 8B parameter language models. Finally, we conjecture the underlying reason for this inductive bias by exploring the connection of stacking to looped models and provide strong supporting empirical analysis.

Large deep learning models have achieved state-of-the-art performance across various natural language processing (NLP) tasks and demonstrated remarkable few-shot learning performance. However, training them is often challenging and resource-intensive. In this paper, we study an efficient approach to train language models using few-shot learners. We show that, by leveraging the fast learning nature of few-shot learners, one can train language models efficiently in a stagewise manner. Our main insight is that stacking a good few-shot learner on a good small language model provides a good initializer for a larger language model. Using this insight and building upon progressive stacking approaches, we develop novel approaches for training such networks in a stagewise manner. Furthermore, we also provide a theoretical framework and accompanying empirical studies to support our insights, thereby creating a theoretical foundation for progressive stacking. Finally, we provide empirical results to demonstrate the effectiveness of our approach in reducing the training time of few-shot learners.

Soil moisture is crucial for agriculture and hydrology, but its accurate prediction is challenging due to inadequate representation of various complex land surface processes and meteorological influences. In this research, we employ the Long Short-Term Memory (LSTM) framework, a specific architecture of deep learning networks that is effective in processing time series data, for predicting soil moisture. We have developed the Next Generation Interactive Soil Moisture Forecasting System to advance skillful soil moisture predictions at sub-seasonal timescales by leveraging advanced analytics and deep learning, with LSTM at its core. We combined the state-of-the-art climate model's (Community Earth System Model Version 2) forecast that incorporates the effects of the large-scale climatic drivers, including sea-surface temperature and atmosphere circulation features into soil water forecast with the LSTM-based Deep Learning model. Our Deep Learning model understands the local forecast biases using the weekly hindcast data from 1999 to 2016. We used this trained LSTM model to test its performance from 2017 to 2021 and enhanced the forecast proficiency and aid in analyzing future soil moisture anomalies, i.e., departure from climatology using data fusion and spatial downscaling. For performance assessment, optimal metrics include Mean Absolute Error (MAE) values near 0 (0-0.6), Root Mean Square Error (RMSE) around 0.5, and Anomaly Correlation Coefficient (ACC) nearing 1. These breakthroughs in system design and modeling facilitate improved soil moisture prediction, benefiting water management and our understanding of land-atmosphere interactions.

Cascades are a classical strategy to enable inference cost to vary adaptively across samples, wherein a sequence of classifiers are invoked in turn. A deferral rule determines whether to invoke the next classifier in the sequence, or to terminate prediction. One simple deferral rule employs the confidence of the current classifier, e.g., based on the maximum predicted softmax probability. Despite being oblivious to the structure of the cascade --- e.g., not modelling the errors of downstream models --- such confidence-based deferral often works remarkably well in practice. In this paper, we seek to better understand the conditions under which confidence-based deferral may fail, and when alternate deferral strategies can perform better. We first present a theoretical characterisation of the optimal deferral rule, which precisely characterises settings under which confidence-based deferral may suffer. We then study post-hoc deferral mechanisms, and demonstrate they can significantly improve upon confidence-based deferral in settings where (i) downstream models are specialists that only work well on a subset of inputs, (ii) samples are subject to label noise, and (iii) there is distribution shift between the train and test set.

Knowledge distillation (KD) has been widely used to improve the test accuracy of a "student" network, by training it to mimic the soft probabilities of a trained "teacher" network. Yet, it has been shown in recent work that, despite being trained to fit the teacher's probabilities, the student may not only significantly deviate from the teacher probabilities, but may also outdo than the teacher in performance. Our work aims to reconcile this seemingly paradoxical observation. Specifically, we characterize the precise nature of the student-teacher deviations, and argue how they _can_ co-occur with better generalization. First, through experiments on image and language data, we identify that these probability deviations correspond to the student systematically _exaggerating_ the confidence levels of the teacher. Next, we theoretically and empirically establish another form of exaggeration in some simple settings: KD exaggerates the implicit bias of gradient descent in converging faster along the top eigendirections of the data. Finally, we tie these two observations together: we demonstrate that the exaggerated bias of KD can simultaneously result in both (a) the exaggeration of confidence and (b) the improved generalization of the student, thus offering a resolution to the apparent paradox. Our analysis brings existing theory and practice closer by considering the role of gradient descent in KD and by demonstrating the exaggerated bias effect in both theoretical and empirical settings.

This paper introduces SOAR: **S**pilling with **O**rthogonality-**A**mplified **R**esiduals, a novel data indexing technique for approximate nearest neighbor (ANN) search. SOAR extends upon previous approaches to ANN search, such as spill trees, that utilize multiple redundant representations while partitioning the data to reduce the probability of missing a nearest neighbor during search. Rather than training and computing these redundant representations independently, however, SOAR uses an *orthogonality-amplified residual* loss, which optimizes each representation to compensate for cases where other representations perform poorly. This drastically improves the overall index quality, resulting in state-of-the-art ANN benchmark performance while maintaining fast indexing times and low memory consumption.

The impressive generalization performance of modern neural networks is attributed in part to their ability to implicitly memorize complex training patterns. Inspired by this, we explore a novel mechanism to improve model generalization via explicit memorization. Specifically, we propose the residual-memorization (ResMem) algorithm, a new method that augments an existing prediction model (e.g., a neural network) by fitting the model's residuals with a nearest-neighbor based regressor. The final prediction is then the sum of the original model and the fitted residual regressor. By construction, ResMem can explicitly memorize the training labels. We start by formulating a stylized linear regression problem and rigorously show that ResMem results in a more favorable test risk over a base linear neural network. Then, we empirically show that ResMem consistently improves the test set generalization of the original prediction model across standard vision and natural language processing benchmarks.

In contrast to SGD, adaptive gradient methods like Adam allow robust training of modern deep networks, especially large language models. However, the use of adaptivity not only comes at the cost of extra memory but also raises the fundamental question: can non-adaptive methods like SGD enjoy similar benefits? In this paper, we provide an affirmative answer to this question by proposing to achieve both robust and memory-efficient training via the following general recipe: (1) modify the architecture and make it scale invariant, (2) train with SGD and weight decay, and optionally (3) clip the global gradient norm proportional to weight norm multiplied by $\sqrt{\frac{2\lambda}{\eta}}$, where $\eta$ is learning rate and $\lambda$ is weight decay. We show that this general approach is robust to rescaling of parameter and loss by proving that its convergence only depends logarithmically on the scale of initialization and loss, whereas the standard SGD might not even converge for many initializations. Following our recipe, we design a scale invariant version of BERT, called SIBERT, which when trained simply by vanilla SGD achieves performance comparable to BERT trained by adaptive methods like Adam on downstream tasks.

Transformer-based models such as BERT have proven successful in information retrieval problem, which seek to identify relevant documents for a given query. There are two broad flavours of such models: cross-attention (CA) models, which learn a joint embedding for the query and document, and dual-encoder (DE) models, which learn separate embeddings for the query and document. Empirically, CA models are often found to be more accurate, which has motivated a series of works seeking to bridge this gap. However, a more fundamental question remains less explored: does this performance gap reflect an inherent limitation in the capacity of DE models, or a limitation in the training of such models? And does such an understanding suggest a principled means of improving DE models? In this paper, we study these questions, with three contributions. First, we establish theoretically that with a sufficiently large embedding dimension, DE models have the capacity to model a broad class of score distributions. Second, we show empirically that on real-world problems, DE models may overfit to spurious correlations in the training set, and thus under-perform on test samples. To mitigate this behaviour, we propose a suitable distillation strategy, and confirm its practical efficacy on the MSMARCO-Passage and Natural Questions benchmarks.

This paper presents a novel nearest neighbor search algorithm achieving TPU (Google Tensor Processing Unit) peak performance, outperforming state-of-the-art GPU algorithms with similar level of recall. The design of the proposed algorithm is motivated by an accurate accelerator performance model that takes into account both the memory and instruction bottlenecks. Our algorithm comes with an analytical guarantee of recall in expectation and does not require maintaining sophisticated index data structure or tuning, making it suitable for applications with frequent updates. Our work is available in the open-source package of Jax and Tensorflow on TPU.

Language models can be augmented with context retriever to incorporate knowledge from large external databases. By leveraging retrieved context, the neural network does not have to memorize the massive amount of world knowledge within its internal parameters, leading to better parameter efficiency, interpretability and modularity. In this paper we examined a simple yet effective architecture for incorporating external context into language models based on decoupled $\texttt{Encoder-Decoder}$ architecture. We showed that such a simple architecture achieves competitive results on auto-regressive language modeling and open domain question answering tasks. We also analyzed the behavior of the proposed model which performs grounded context transfer. Finally we discussed the computational implications of such retrieval augmented models.

Many practical settings allow a learner to defer predictions to one or more costly experts. For example, the learning to defer paradigm allows a learner to defer to a human expert, at some monetary cost. Similarly, the adaptive inference paradigm allows a base model to defer to one or more large models, at some computational cost. The goal in these settings is to learn classification and deferral mechanisms to optimise a suitable accuracy-cost tradeoff. To achieve this, a central issue studied in prior work is the design of a coherent loss function for both mechanisms. In this work, we demonstrate that existing losses have two subtle limitations: they can encourage underfitting when there is a high cost of deferring, and the deferral function can have a weak dependence on the base model predictions. To resolve these issues, we propose a post-hoc training scheme: we train a deferral function on top of a base model, with the objective of predicting to defer when the base model's error probability exceeds the cost of the expert model. This may be viewed as applying a partial surrogate to the ideal deferral loss, which can lead to a tighter approximation and thus better performance. Empirically, we verify the efficacy of post-hoc training on benchmarks for learning to defer and adaptive inference.

Knowledge distillation is an approach to improve the performance of a student model by using the knowledge of a complex teacher. Despite its success in several deep learning applications, the study of distillation is mostly confined to classification settings. In particular, the use of distillation in top-k ranking settings, where the goal is to rank k most relevant items correctly, remains largely unexplored. In this paper, we study such ranking problems through the lens of distillation. We present a distillation framework for top-k ranking and draw connections with the existing ranking methods. The core idea of this framework is to preserve the ranking at the top by matching the order of items of student and teacher, while penalizing large scores for items ranked low by the teacher. Building on this, we develop a novel distillation approach, RankDistil, specifically catered towards ranking problems with a large number of items to rank, and establish statistical basis for the method. Finally, we conduct experiments which demonstrate that RankDistil yields benefits over commonly used baselines for ranking problems.

Knowledge distillation is a technique for improving a “student” model by replacing its one-hot training labels with a label distribution obtained from a “teacher” model. Despite its broad success, several basic questions — e.g., Why does distillation help? Why do more accurate teachers not necessarily distill better? — have received limited formal study. In this paper, we present a statistical perspective on distillation which provides an answer to these questions. Our core observation is that a “Bayes teacher” providing the true class-probabilities can lower the variance of the student objective, and thus improve performance. We then establish a bias-variance tradeoff that quantifies the value of teachers that approximate the Bayes class-probabilities. This provides a formal criterion as to what constitutes a “good” teacher, namely, the quality of its probability estimates. Finally, we illustrate how our statistical perspective facilitates novel applications of distillation to bipartite ranking and multiclass retrieval.

Negative sampling schemes enable efficient training given a large number of classes, by offering a means to approximate a computationally expensive loss function that takes all labels into account. In this paper, we present a new connection between these schemes and loss modification techniques for countering label imbalance. We show that different negative sampling schemes implicitly trade-off performance on dominant versus rare labels. Further, we provide a unified means to explicitly tackle both sampling bias, arising from working with a subset of all labels, and labeling bias, which is inherent to the data due to label imbalance. We empirically verify our findings on long-tail classification and retrieval benchmarks.

The ability to train complex and highly effective models often requires an abundance of training data, which can easily become a bottleneck in cost, time, and computational resources. Batch active learning, which adaptively issues batched queries to a labeling oracle, is a common approach for addressing this problem. The practical benefits of batch sampling come with the downside of less adaptivity and the risk of sampling redundant examples within a batch -- a risk that grows with the batch size. In this work, we analyze an efficient active learning algorithm, which focuses on the large batch setting. In particular, we show that our sampling method, which combines notions of uncertainty and diversity, easily scales to batch sizes (100K-1M) several orders of magnitude larger than used in previous studies and provides significant improvements in model training efficiency compared to recent baselines. Finally, we provide an initial theoretical analysis, proving label complexity guarantees for a related sampling method, which we show is approximately equivalent to our sampling method in specific settings.

Factorized models, such as two tower neural network models, are widely used for scoring (query, document) pairs in information retrieval tasks. These models are typically trained by optimizing the model parameters to score relevant positive" pairs higher than the irrelevantnegative" ones. While a large set of negatives typically improves the model performance, limited computation and memory budgets place constraints on the number of negatives used during training. In this paper, we develop a novel negative sampling technique for accelerating training with softmax cross-entropy loss. By using cached (possibly stale) item embeddings, our technique enables training with a large pool of negatives with reduced memory and computation. We also develop a streaming variant of our algorithm geared towards very large datasets. Furthermore, we establish a theoretical basis for our approach by showing that updating a very small fraction of the cache at each iteration can still ensure fast convergence. Finally, we experimentally validate our approach and show that it is efficient and compares favorably with more complex, state-of-the-art approaches.

We characterize the minimum noise amplitude and power for noise-adding mechanisms in (epsilon, delta)-differential privacy for single real-valued query function. We derive new lower bounds using the duality of linear programming, and new upper bounds by analyzing a special class of (epsilon, delta)-differentially private mechanisms, the truncated Laplacian mechanisms. We show that the multiplicative gap of the lower bounds and upper bounds goes to zero in various high privacy regimes, proving the tightness of the lower and upper bounds. In particular, our results close the previous constant multiplicative gap in the discrete setting. Numeric experiments show the improvement of the truncated Laplacian mechanism over the optimal Gaussian mechanism in all privacy regimes.

Neural Ordinary Differential Equation (Neural ODE) has been proposed as a continuous approximation to the ResNet architecture. Some commonly used regularization mechanisms in discrete neural networks (e.g., dropout, Gaussian noise) are missing in current Neural ODE networks. In this paper, we propose a new continuous neural network framework called Neural Stochastic Differential Equation (Neural SDE), which naturally incorporates various commonly used regularization mechanisms based on random noise injection. For regularization purposes, our framework includes multiple types of noise patterns, such as dropout, additive, and multiplicative noise, which are common in plain neural networks. We provide some theoretical analyses explaining the improved robustness of our models against input perturbations. Furthermore, we demonstrate that the Neural SDE network can achieve better generalization than the Neural ODE and is more resistant to adversarial and non-adversarial input perturbations.

Michal Lukasik, Himanshu Jain, Aditya Menon, Seungyeon Kim, Srinadh Bhojanapalli, Felix Yu, Sanjiv Kumar. Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP). 2020.

Attention based Transformer architecture has enabled significant advances in the field of natural language processing. In addition to new pre-training techniques, recent improvements crucially rely on working with a relatively larger embedding dimension for tokens. Unfortunately, this leads to models that are prohibitively large to be employed in the downstream tasks. In this paper we identify one of the important factors contributing to the large embedding size requirement. In particular, our analysis highlights that the scaling between the number of heads and the size of each head in the current architecture gives rise to a low-rank bottleneck in attention heads, causing this limitation. We further validate this in our experiments. As a solution we propose to set the head size of an attention unit to input sequence length, and independent of the number of heads, resulting in multi-head attention layers with provably more expressive power. We empirically show that this allows us to train models with a relatively smaller embedding dimension and with better performance scaling.

Quantization based techniques are the current state-of-the-art for scaling maximum inner product search to massive databases. Traditional approaches to quantization aim to minimize the reconstruction error of the database points. Based on the observation that for a given query, the database points that have the largest inner products are more relevant, we develop a family of anisotropic quantization loss functions. Under natural statistical assumptions, we show that quantization with these loss functions leads to a new variant of vector quantization that more greatly penalizes the parallel component of a datapoint’s residual relative to its orthogonal component. The proposed approach, whose implementation is open-source, achieves state-of-the-art results on the public benchmarks available at ann-benchmarks.com.

Label smoothing is commonly used in training deep learning models, wherein one-hot training labels are mixed with uniform label vectors. Empirically, smoothing has been shown to improve both predictive performance and model calibration. In this paper, we study whether label smoothing is also effective as a means of coping with label noise. While label smoothing apparently amplifies this problem — being equivalent to injecting symmetric noise to the labels — we show how it relates to a general family of loss-correction techniques from the label noise literature. Building on this connection, we show that label smoothing is competitive with loss-correction under label noise. Further, we show that when distilling models from noisy data, label smoothing of the teacher is beneficial; this is in contrast to recent findings for noise-free problems, and sheds further light on settings where label smoothing is beneficial.

We consider learning a multi-class classification model in the federated setting, where each user has access to the positive data associated with only a single class. As a result, during each federated learning round, the users need to locally update the classifier without having access to the features and the model parameters for the negative classes. Thus, naively employing conventional decentralized learning such as distributed SGD or Federated Averaging may lead to trivial or extremely poor classifiers. In particular, for embedding based classifiers, all the class embeddings might collapse to a single point. To address this problem, we propose a generic framework for training with only positive labels, namely Federated Averaging with Spreadout (FedAwS), where the server imposes a geometric regularizer after each round to encourage classes to be spreadout in the embedding space. We show, both theoretically and empirically, that FedAwS can almost match the performance of conventional learning where users have access to negative labels. We further extend the proposed method to settings with large output spaces.

Recently, there has been a surge of interest in representation learning in hyperbolic spaces, driven by their ability to represent hierarchical data with significantly fewer dimensions than standard Euclidean spaces. However, the viability and benefits of hyperbolic spaces for downstream machine learning tasks have received less attention. In this paper, we present, to our knowledge, the first theoretical guarantees for learning a classifier in hyperbolic rather than Euclidean space. Specifically, we consider the problem of learning a large-margin classifier for data possessing a hierarchical structure. Our first contribution is a hyperbolic perceptron algorithm, which provably converges to a separating hyperplane. We then provide an algorithm to efficiently learn a large-margin hyperplane, relying on the careful injection of adversarial examples. Finally, we prove that for hierarchical data that embeds well into hyperbolic space, the low embedding dimension ensures superior guarantees when learning the classifier directly in hyperbolic space.

Multi-stage training and knowledge transfer, from a large-scale pretraining task to various finetuning tasks, have revolutionized natural language processing and computer vision resulting in state-of-the-art performance improvements. In this paper, we develop a multi-stage influence function score to track predictions from a finetuned model all the way back to the pretraining data. With this score, we can identify the pretraining examples in the pretraining task that contribute most to a prediction in the finetuning task. The proposed multi-stage influence function generalizes the original influence function for a single model in (Koh &Liang, 2017), thereby enabling influence computation through both pretrained and finetuned models. We study two different scenarios with the pretrained embedding fixed or updated in the finetuning tasks. We test our proposed method in various experiments to show its effectiveness and potential applications.

Much of the literature on differential privacy focuses on item-level privacy, where loosely speaking, the goal is to provide privacy per item or training example. However, recently many practical applications such as federated learning require preserving privacy for all items of a single user, which is much harder to achieve. Therefore understanding the theoretical limit of user-level privacy becomes crucial. We study the fundamental problem of learning discrete distributions over $k$ symbols with user-level differential privacy. If each user has $m$ samples, we show that straightforward applications of Laplace or Gaussian mechanisms require the number of users to be $\mathcal{O}(k/(m\alpha^2) + k/\epsilon\alpha)$ to achieve an $\ell_1$ distance of $\alpha$ between the true and estimated distributions, with the privacy-induced penalty $k/\epsilon\alpha$ independent of the number of samples per user $m$. Moreover, we show that any mechanism that only operates on the final aggregate should require a user complexity of the same order. We then propose a mechanism such that the number of users scales as $\tilde{\mathcal{O}}(k/(m\alpha^2) + k/\sqrt{m}\epsilon\alpha)$ and further show that it is nearly-optimal under certain regimes. Thus the privacy penalty is $\tilde{\Theta}(\sqrt{m})$ times smaller compared to the standard mechanisms. We also propose general techniques for obtaining lower bounds on restricted differentially private estimators and a lower bound on the total variation between binomial distributions, both of which might be of independent interest.

Recently, Transformer networks have redefined the state of the art in many NLP tasks. However, these models suffer from quadratic computational cost in the input sequence length $n$ to compute pairwise attention in each layer. This has prompted recent research into sparse Transformers that sparsify the connections in the attention layers. While empirically promising for long sequences, fundamental questions remain unanswered: Can sparse Transformers approximate any arbitrary sequence-to-sequence function, similar to their dense counterparts? How does the sparsity pattern and the sparsity level affect their performance? In this paper, we address these questions and provide a unifying framework that captures existing sparse attention models. We propose sufficient conditions under which we prove that a sparse attention model can universally approximate any sequence-to-sequence function. Surprisingly, our results show that sparse Transformers with only $O(n)$ connections per attention layer can approximate the same function class as the dense model with $n^2$ connections. Lastly, we present experiments comparing different patterns/levels of sparsity on standard NLP tasks.

While stochastic gradient descent (SGD) is still the de facto algorithm in deep learning, adaptive methods like Clipped SGD/Adam have been observed to outperform SGD across important tasks, such as attention models. The settings under which SGD performs poorly in comparison to adaptive methods are not well understood yet. In this paper, we provide empirical and theoretical evidence that a heavy-tailed distribution of the noise in stochastic gradients is one cause of SGD's poor performance. We provide the first tight upper and lower convergence bounds for adaptive gradient methods under heavy-tailed noise. Further, we demonstrate how gradient clipping plays a key role in addressing heavy-tailed gradient noise. Subsequently, we show how clipping can be applied in practice by developing an adaptive coordinate-wise clipping algorithm (ACClip) and demonstrate its superior performance on BERT pretraining and finetuning tasks.

Random Fourier features are a powerful framework to approximate shift invariant kernels with Monte Carlo integration, which has drawn considerable interest in scaling up kernel-based learning, dimensionality reduction, and information retrieval. In the literature, many sampling schemes have been proposed to improve the approximation performance. However, an interesting theoretic and algorithmic challenge still remains, i.e., how to optimize the design of random Fourier features to achieve good kernel approximation on any input data using a low spectral sampling rate? In this paper, we propose to compute more adaptive random Fourier features with optimized spectral samples (wj’s) and feature weights (pj’s). The learning scheme not only significantly reduces the spectral sampling rate needed for accurate kernel approximation, but also allows joint optimization with any supervised learning framework. We establish generalization bounds using Rademacher complexity, and demonstrate advantages over previous methods. Moreover, our experiments show that the empirical kernel approximation provides effective regularization for supervised learning.

We derive the optimal $(0, \delta)$-differentially private query-output independent noise-adding mechanism for single real-valued query function under a general cost-minimization framework. Under a mild technical condition, we show that the optimal noise probability distribution is a uniform distribution with a probability mass at the origin. We explicitly derive the optimal noise distribution for general $\ell^p$ cost functions, including $\ell^1$ (for noise magnitude) and $\ell^2$ (for noise power) cost functions, and show that the probability concentration on the origin occurs when $\delta > \frac{p}{p+1}$. Our result demonstrates an improvement over the existing Gaussian mechanisms by a factor of two and three for $(0,\delta)$-differential privacy in the high privacy regime in the context of minimizing the noise magnitude and noise power, and the gain is more pronounced in the low privacy regime. Our result is consistent with the existing result for $(0,\delta)$-differential privacy in the discrete setting, and identifies a probability concentration phenomenon in the continuous setting.

We consider the problem of retrieving the most relevant labels for a given input when the size of the output space is very large. Retrieval methods are modeled as set-valued classifiers which output a small set of classes for each input, and a mistake is made if the label is not in the output set. Despite its practical importance, a statistically principled, yet practical solution to this problem is largely missing. To this end, we first define a family of surrogate losses and show that they are calibrated and convex under certain conditions on the loss parameters and data distribution, thereby establishing a statistical and analytical basis for using these losses. Furthermore, we identify a particularly intuitive class of loss functions in the aforementioned family and show that they are amenable to practical implementation in the large output space setting (i.e. computation is possible without evaluating scores of all labels) by developing a technique called Stochastic Negative Mining. We also provide generalization error bounds for the losses in the family. Finally, we conduct experiments which demonstrate that Stochastic Negative Mining yields benefits over commonly used negative sampling approaches.

Adaptive methods such as Adam and RMSProp are widely used in deep learning but are not well understood. In this paper, we seek a crisp, clean and precise characterization of their behavior in nonconvex settings. To this end, we first provide a novel view of adaptive methods as preconditioned SGD, where the preconditioner is estimated in an online manner. By studying the preconditioner on its own, we elucidate its purpose: it rescales the stochastic gradient noise to be isotropic near stationary points, which helps escape saddle points. Furthermore, we show that adaptive methods can efficiently estimate the aforementioned preconditioner. By gluing together these two components, we provide the first (to our knowledge) second-order convergence result for any adaptive method. The key insight from our analysis is that, compared to SGD, adaptive methods escape saddle points faster, and can converge faster overall to second-order stationary points.

Linear encoding of sparse vectors is widely popular, but is commonly data-independent – missing any possible extra (but a priori unknown) structure beyond sparsity. In this paper we present a new method to learn linear encoders that adapt to data, while still performing well with the widely used $\ell_1$ decoder. The convex $\ell_1$ decoder prevents gradient propagation as needed in standard gradient-based training. Our method is based on the insight that unrolling the convex decoder into $T$ projected subgradient steps can address this issue. Our method can be seen as a data-driven way to learn a compressed sensing measurement matrix. We compare the empirical performance of 10 algorithms over 6 sparse datasets (3 synthetic and 3 real). Our experiments show that there is indeed additional structure beyond sparsity in the real datasets; our method is able to discover it and exploit it to create excellent reconstructions with fewer measurements (by a factor of 1.1-3x) compared to the previous state-of-the-art methods. We illustrate an application of our method in learning label embeddings for extreme multi-label classification, and empirically show that our method is able to match or outperform the precision scores of SLEEC, which is one of the state-of-the-art embedding-based approaches.

In extreme classification settings, embedding-based neural network models are currently not competitive with sparse linear and tree-based methods in terms of accuracy. Most prior works attribute this poor performance to the low-dimensional bottleneck in embedding-based methods. In this paper, we demonstrate that theoretically there is no limitation to using low-dimensional embedding-based methods, and provide experimental evidence that overfitting is the root cause of the poor performance of embedding-based methods. These findings motivate us to investigate novel data augmentation and regularization techniques to mitigate overfitting. To this end, we propose GLaS, a new regularizer for embedding-based neural network approaches. It is a natural generalization from the graph Laplacian and spread-out regularizers, and empirically it addresses the drawback of each regularizer alone when applied to the extreme classification setup. With the proposed techniques, we attain or improve upon the state-of-the-art on most widely tested public extreme classification datasets with hundreds of thousands of labels.

The computational cost of training with softmax cross entropy loss grows linearly with the number of classes. For the settings where a large number of classes are involved, a common method to speed up training is to sample a subset of classes and utilize an estimate of the loss gradient based on these classes, known as the sampled softmax method. However, the sampled softmax provides a biased estimate of the gradient unless the samples are drawn from the exact softmax distribution, which is again expensive to compute. Therefore, a widely employed practical approach involves sampling from a simpler distribution in the hope of approximating the exact softmax distribution. In this paper, we develop the first theoretical understanding of the role that different sampling distributions play in determining the quality of sampled softmax. Motivated by our analysis and the work on kernel-based sampling, we propose the Random Fourier Softmax (RF-softmax) method that utilizes the powerful Random Fourier Features to enable more efficient and accurate sampling from an approximate softmax distribution. We show that RF-softmax leads to low bias in estimation in terms of both the full softmax distribution and the full softmax gradient. Furthermore, the cost of RF-softmax scales only logarithmically with the number of classes.

For problems with large output spaces, evaluation of the loss function and its gradient are expensive, typically taking linear time in the size of the output space. Recently, methods have been developed to speed up learning via efficient data structures for Nearest-Neighbor Search (NNS) or Maximum Inner-Product Search (MIPS). However, the performance of such data structures typically degrades in high dimensions. In this work, we propose a novel technique to reduce the intractable high dimensional search problem to several much more tractable lower dimensional ones via dual decomposition of the loss function. At the same time, we demonstrate guaranteed convergence to the original loss via a greedy message passing procedure. In our experiments on multiclass and multilabel classification with hundreds of thousands of classes, as well as training skip-gram word embeddings with a vocabulary size of half a million, our technique consistently improves the accuracy of search-based gradient approximation methods and outperforms sampling-based gradient approximation methods by a large margin.

Distributed stochastic gradient descent is an important subroutine in distributed learning. A setting of particular interest is when the clients are mobile devices, where two important concerns are communication efficiency and the privacy of the clients. Several recent works have focused on reducing the communication cost or introducing privacy guarantees, but none of the proposed communication efficient methods are known to be privacy preserving and none of the known privacy mechanisms are known to be communication efficient. To this end, we study algorithms that achieve both communication efficiency and differential privacy. For $d$ variables and $n \approx d$ clients, the proposed method uses $\cO(\log \log(nd))$ bits of communication per client per coordinate and ensures constant privacy. We also improve previous analysis of the \emph{Binomial mechanism} showing that it achieves nearly the same utility as the Gaussian mechanism, while requiring fewer representation bits, which can be of independent interest.

Adaptive gradient methods that rely on scaling gradients down by the square root of exponential moving averages of past squared gradients, such RMSProp, Adam, Adadelta have found wide application in optimizing the nonconvex problems that arise in deep learning. However, it has been recently demonstrated that such methods can fail to converge even in simple convex optimization settings. In this work, we provide a new analysis of such methods applied to nonconvex stochastic optimization problems, characterizing the effect of increasing minibatch size. Our analysis shows that under this scenario such methods do converge to stationarity up to the statistical limit of variance in the stochastic gradients (scaled by a constant factor). In particular, our result implies that increasing minibatch sizes enables convergence, thus providing a way to circumvent the non-convergence issues. Furthermore, we provide a new adaptive optimization algorithm, Yogi, which controls the increase in effective learning rate, leading to even better performance with similar theoretical guarantees on convergence. Extensive experiments show that Yogi with very little hyperparameter tuning outperforms methods such as Adam in several challenging machine learning tasks.

In this work, we propose a new technique for \emphfast k-nearest neighbor (k-NN) classification in which the original database is represented via a small set of learned binary prototypes. The training phase simultaneously learns a hash function which maps the data points to binary codes, and a set of representative binary prototypes. In the prediction phase, we first hash the query into a binary code and then do the k-NN classification using the binary prototypes as the database. Our approach speeds up k-NN classification in two aspects. First, we compress the database into a smaller set of prototypes such that k-NN search only goes through a smaller set rather than the whole dataset. Second, we reduce the original space to a compact binary embedding, where the Hamming distance between two binary codes is very efficient to compute. We propose a formulation to learn the hash function and prototypes such that the classification error is minimized. We also provide a novel theoretical analysis of the proposed technique in terms of Bayes error consistency. Empirically, our method is much faster than the state-of-the-art k-NN compression methods with comparable accuracy.

We propose a simple, yet powerful regularization technique that can be used to significantly improve both the pairwise and triplet losses in learning local feature descriptors. The idea is that in order to fully utilize the expressive power of the descriptor space, good local feature descriptors should be sufficiently “spread-out” over the space. In this work, we propose a regularization term to maximize the spread in feature descriptor inspired by the property of uniform distribution. We show that the proposed regularization with triplet loss outperforms existing Euclidean distance based descriptor learning techniques by a large margin. As an extension, the proposed regularization technique can also be used to improve image-level deep feature embedding.

Learning-based binary hashing has become a powerful paradigm for fast search and retrieval in massive databases. However, due to the requirement of discrete outputs for the hash functions, learning such functions is known to be very challenging. In addition, the objective functions adopted by existing hashing techniques are mostly chosen heuristically. In this paper, we propose a novel generative approach to learn hash functions through Minimum Description Length principle such that the learned hash codes maximally compress the dataset and can also be used to regenerate the inputs. We also develop an efficient learning algorithm based on the stochastic distributional gradient, which avoids the notorious difficulty caused by binary output constraints, to jointly optimize the parameters of the hash function and the associated generative model. Extensive experiments on a variety of large-scale datasets show that the proposed method achieves better retrieval results than the existing state-of-the-art methods.

Motivated by the need for distributed learning and optimization algorithms with low communication cost, we study communication efficient algorithms for distributed mean estimation. Unlike previous works, we make no probabilistic assumptions on the data. We first show that for $d$ dimensional data with $n$ clients, a naive stochastic rounding approach yields a mean squared error (MSE) of $\Theta(d/n)$ and uses a constant number of bits per dimension per client. We then extend this naive algorithm in two ways: we show that applying a structured random rotation before quantization reduces the error to $\mathcal{O}((\log d)/n)$ and a better coding strategy further reduces the error to $\mathcal{O}(1/n)$. We also show that the latter coding strategy is optimal up to a constant in the minimax sense i.e., it achieves the best MSE for a given communication cost. We finally demonstrate the practicality of our algorithms by applying them to distributed Lloyd’s algorithm for k-means and power iteration for PCA.

We propose a multiscale quantization approach for fast similarity search on large, high-dimensional datasets. The key insight of the approach is that quantization methods, in particular product quantization, perform poorly when there is large variance in the norms of the data points. This is a common scenario for real- world datasets, especially when doing product quantization of residuals obtained from coarse vector quantization. To address this issue, we propose a multiscale formulation where we learn a separate scalar quantizer of the residual norm scales. All parameters are learned jointly in a stochastic gradient descent framework to minimize the overall quantization error. We provide theoretical motivation for the proposed technique and conduct comprehensive experiments on two large-scale public datasets, demonstrating substantial improvements in recall over existing state-of-the-art methods.

We propose a quantization based approach for fast approximate Maximum Inner Product Search (MIPS). Each database vector is quantized in multiple subspaces via a set of codebooks, learned directly by minimizing the inner product quantization error. Then, the inner product of a query to a database vector is approximated as the sum of inner products with the subspace quantizers. Different from recently proposed LSH approaches to MIPS, the database vectors and queries do not need to be augmented in a higher dimensional feature space. We also provide a theoretical analysis of the proposed approach, consisting of the concentration results under mild assumptions. Furthermore, if a small set of held-out queries is given at the training time, we propose a modified codebook learning procedure which further improves the accuracy. Experimental results on a variety of datasets including those arising from deep neural networks show that the proposed approach significantly outperforms the existing state-of-the-art.

We consider the hashing mechanism for constructing binary embeddings, that involves pseudo-random projections followed by nonlinear (sign function) mappings. The pseudo-random projection is described by a matrix, where not all entries are independent random variables but instead a fixed “budget of randomness” is distributed across the matrix. Such matrices can be edfficiently stored in sub-quadratic or even linear space, provide reduction in randomness usage (i.e. number of required random values), and very often lead to computational speed ups. We prove several theoretical results showing that projections via various structured matrices followed by nonlinear mappings accurately preserve the angular distance between input high-dimensional vectors. To the best of our knowledge, these results are the first that give theoretical ground for the use of general structured matrices in the nonlinear setting. We empirically verify our theoretical findings and show the dependence of learning via structured hashed projections on the performance of neural network as well as nearest neighbor classifier.

We present an intriguing discovery related to Random Fourier Features: replacing multiplication by a random Gaussian matrix with multiplication by a properly scaled random orthogonal matrix significantly decreases kernel approximation error. We call this technique Orthogonal Random Features (ORF), and provide theoretical and empirical justification for its effectiveness. Motivated by the discovery, we further propose Structured Orthogonal Random Features (SORF), which uses a class of structured discrete orthogonal matrices to speed up the computation. The method reduces the time cost from $\mathcal{O}(d^2)$ to $\mathcal{O}(d \log d)$, where $d$ is the data dimensionality, with almost no compromise in kernel approximation quality compared to ORF. Experiments on several datasets verify the effectiveness of ORF and SORF over the existing methods. We also provide discussions on using the same type of discrete orthogonal structure for a broader range of kernels and applications.

Hydroclimatic research requires highly intensive resources in terms of computation and data to perform simulations. Setting up complex experiment environment and configurations to submit jobs in computational clusters as well as managing user’s limited storage spaces by transferring big size data into the secondary storage are complicated and time-consuming. As a possible answer to address such issues in hydroclimatic research, new technologies, software-defined storage and containers have been introduced. When the two technologies are combined to support hydroclimatic simulations, we discuss how the software-defined storage data infrastructure strengthens containers in terms of flexibility of data handling and storage scalability.

The explosive growth in Big Data has attracted much attention in designing efficient indexing and search methods recently. In many critical applications such as large-scale search and pattern matching, finding the nearest neighbors to a query is a fundamental research problem. However, the straightforward solution using exhaustive comparison is infeasible due to the prohibitive computational complexity and memory requirement. In response, approximate nearest neighbor (ANN) search based on hashing techniques has become popular due to its promising performance in both efficiency and accuracy. Prior randomized hashing methods, e.g., locality-sensitive hashing (LSH), explore data-independent hash functions with random projections or permutations. Although having elegant theoretic guarantees on the search quality in certain metric spaces, performance of randomized hashing has been shown insufficient in many real-world applications. As a remedy, new approaches incorporating data-driven learning methods in development of advanced hash functions have emerged. Such learning-to-hash methods exploit information such as data distributions or class labels when optimizing the hash codes or functions. Importantly, the learned hash codes are able to preserve the proximity of neighboring data in the original feature spaces in the hash code spaces. The goal of this paper is to provide readers with systematic understanding of insights, pros, and cons of the emerging techniques. We provide a comprehensive survey of the learning-to-hash framework and representative techniques of various types, including unsupervised, semisupervised, and supervised. In addition, we also summarize recent hashing approaches utilizing the deep learning models. Finally, we discuss the future direction and trends of research in this area.

This paper describes a large scale exemplar-based acoustic modeling approach for large vocabulary continuous speech recognition. We construct an index of labeled training frames using high-level features extracted from the bottleneck layer of a deep neural network as indexing features. At recognition time, each test frame is turned into a query and a set of k-nearest neighbor frames is retrieved from the index. This set is further filtered using majority voting and the remaining frames are used to derive an estimate of the context-dependent state posteriors of the query, which can then be used for recognition. Using an approximate nearest neighbor search approach based on asymmetric hashing, we are able to construct an index on over 25,000 hours of training data. We present both frame classification and recognition experiments on a Voice Search task.

We explore the redundancy of parameters in deep neural networks by replacing the conventional linear projection in fully-connected layers with the circulant projection. The circulant structure substantially reduces memory footprint and enables the use of the Fast Fourier Transform to speed up the computation. Considering a fully-connected neural network layer with d input nodes, and d output nodes, this method improves the time complexity from O(d2) to O(dlogd) and space complexity from O(d2) to O(d). The space savings are particularly important for modern deep convolutional neural network architectures, where fully-connected layers typically contain more than 90% of the network parameters. We further show that the gradient computation and optimization of the circulant projections can be performed very efficiently. Our experiments on three standard datasets show that the proposed approach achieves this significant gain in storage and efficiency with minimal increase in error rate compared to neural networks with unstructured projections.

We propose a family of structured matrices to speed up orthogonal projections for high-dimensional data commonly seen in computer vision applications. In this, a structured matrix is formed by the Kronecker product of a series of smaller orthogonal matrices. This achieves O(dlogd) computational complexity and O(logd) space complexity for d-dimensional data, a drastic improvement over the standard unstructured projections whose computational and space complexities are both O(d^2). The proposed structured matrices are applicable to a number of application domains, and are faster and more compact than other structured matrices used in the past. We also introduce an efficient learning procedure for optimizing such matrices in a data dependent fashion. We demonstrate the significant advantages of the proposed approach in solving the approximate nearest neighbor (ANN) image search problem with both binary embedding and quantization. We find that the orthogonality plays a very important role in solving ANN problem, since the random orthogonal Kronecker projection has already provided promising performance. Comprehensive experiments show that the proposed approach can achieve similar or better accuracy as the existing state-of-the-art but with significantly less time and memory.

Compact explicit feature maps provide a practical framework to scale kernel methods to large-scale learning, but deriving such maps for many types of kernels remains a challenging open problem. Among the commonly used kernels for nonlinear classification are polynomial kernels, for which low approximation error has thus far necessitated explicit feature maps of large dimensionality, especially for higher-order polynomials. Meanwhile, because polynomial kernels are unbounded, they are frequently applied to data that has been normalized to unit l2 norm. The question we address in this work is: if we know a priori that data is so normalized, can we devise a more compact map? We show that a putative affirmative answer to this question based on Random Fourier Features is impossible in this setting, and introduce a new approximation paradigm, Spherical Random Fourier (SRF) features, which circumvents these issues and delivers a compact approximation to polynomial kernels for data on the unit sphere. Compared to prior work, SRF features are less rank-deficient, more compact, and achieve better kernel approximation, especially for higher-order polynomials. The resulting predictions have lower variance and typically yield better classification accuracy.

We consider the task of building compact deep learning pipelines suitable for deploymenton storage and power constrained mobile devices. We propose a uni-fied framework to learn a broad family of structured parameter matrices that arecharacterized by the notion of low displacement rank. Our structured transformsadmit fast function and gradient evaluation, and span a rich range of parametersharing configurations whose statistical modeling capacity can be explicitly tunedalong a continuum from structured to unstructured. Experimental results showthat these transforms can significantly accelerate inference and forward/backwardpasses during training, and offer superior accuracy-compactness-speed tradeoffsin comparison to a number of existing techniques. In keyword spotting applicationsin mobile speech recognition, our methods are much more effective thanstandard linear low-rank bottleneck layers and nearly retain the performance ofstate of the art models, while providing more than 3.5-fold compression.

Binary embedding of high-dimensional data requires long codes to preserve the discriminative power of the input space. Traditional binary coding methods often suffer from very high computation and storage costs in such a scenario. To address this problem, we propose Circulant Binary Embedding (CBE) which generates binary codes by projecting the data with a circulant matrix. The circulant structure enables the use of Fast Fourier Transformation to speed up the computation. Compared to methods that use unstructured matrices, the proposed method improves the time complexity from \mathcalO(d^2) to \mathcalO(d\logd), and the space complexity from \mathcalO(d^2) to \mathcalO(d) where d is the input dimensionality. We also propose a novel time-frequency alternating optimization to learn data-dependent circulant projections, which alternatively minimizes the objective in original and Fourier domains. We show by extensive experiments that the proposed approach gives much better performance than the state-of-the-art approaches for fixed time, and provides much faster computation with no performance degradation for fixed number of bits.

Recent advances in visual recognition indicate that to achieve good retrieval and classification accuracy on large-scale datasets like Image Net, extremely high-dimensional visual descriptors, e.g., Fisher Vectors, are needed. We present a novel method for converting such descriptors to compact similarity-preserving binary codes that exploits their natural matrix structure to reduce their dimensionality using compact bilinear projections instead of a single large projection matrix. This method achieves comparable retrieval and classification accuracy to the original descriptors and to the state-of-the-art Product Quantization approach while having orders of magnitude faster code generation time and smaller memory footprint.

We study the problem of learning with label proportions in which the training data is provided in groups and only the proportion of each class in each group is known. We propose a new method called proportion-SVM, or \proptoSVM, which explicitly models the latent unknown instance labels together with the known group label proportions in a large-margin framework. Unlike the existing works, our approach avoids making restrictive assumptions about the data. The \proptoSVM model leads to a non-convex integer programming problem. In order to solve it efficiently, we propose two algorithms: one based on simple alternating optimization and the other based on a convex relaxation. Extensive experiments on standard datasets show that \proptoSVM outperforms the state-of-the-art, especially for larger group sizes.

This paper focuses on the problem of learning binary embeddings for efficient retrieval of high-dimensional non-negative data. Such data typically arises in a large number of vision and text applications where counts or frequencies are used as features. Also, cosine distance is commonly used as a measure of dissimilarity between such vectors. In this work, we introduce a novel spherical quantization scheme to generate binary embedding of such data and analyze its properties. The number of quantization landmarks in this scheme grows exponentially with data dimensionality resulting in low-distortion quantization. We propose a very efficient method for computing the binary embedding using such large number of landmarks. Further, a linear transformation is learned to minimize the quantization error by adapting the method to the input data resulting in improved embedding. Experiments on image and text retrieval applications show superior performance of the proposed method over other existing state-of-the-art methods.

Hashing-based approximate nearest neighbor (ANN) search in huge databases has become popular due to its computational and memory efficiency. The popular hashing methods, e.g., Locality Sensitive Hashing and Spectral Hashing, construct hash functions based on random or principal projections. The resulting hashes are either not very accurate or are inefficient. Moreover, these methods are designed for a given metric similarity. On the contrary, semantic similarity is usually given in terms of pairwise labels of samples. There exist supervised hashing methods that can handle such semantic similarity, but they are prone to overfitting when labeled data are small or noisy. In this work, we propose a semi-supervised hashing (SSH) framework that minimizes empirical error over the labeled set and an information theoretic regularizer over both labeled and unlabeled sets. Based on this framework, we present three different semi-supervised hashing methods, including orthogonal hashing, nonorthogonal hashing, and sequential hashing. Particularly, the sequential hashing method generates robust codes in which each hash function is designed to correct the errors made by the previous ones. We further show that the sequential learning paradigm can be extended to unsupervised domains where no labeled pairs are available. Extensive experiments on four large datasets (up to 80 million samples) demonstrate the superior performance of the proposed SSH methods over state-of-the-art supervised and unsupervised hashing techniques.

Automatic categorization of videos in a Web-scale unconstrained collection such as YouTube is a challenging task. A key issue is how to build an effective training set in the presence of missing, sparse or noisy labels. We propose to achieve this by first manually creating a small labeled set and then extending it using additional sources such as related videos, searched videos, and text-based webpages. The data from such disparate sources has different properties and labeling quality, and thus fusing them in a coherent fashion is another practical challenge. We propose a fusion framework in which each data source is first combined with the manually-labeled set independently. Then, using the hierarchical taxonomy of the categories, a Conditional Random Field (CRF) based fusion strategy is designed. Based on the final fused classifier, category labels are predicted for the new videos. Extensive experiments on about 80K videos from 29 most frequent categories in YouTube show the effectiveness of the proposed method for categorizing large-scale wild Web videos.

This paper addresses the problem of approximate singular value decomposition of large dense matrices that arises naturally in many machine learning applications. We discuss two recently introduced sampling-based spectral decomposition techniques: the Nyström and the Column-sampling methods. We present a theoretical comparison between the two methods and provide novel insights regarding their suitability for various applications. We then provide experimental results motivated by this theory. Finally, we propose an efficient adaptive sampling technique to select informative columns from the original matrix. This novel technique outperforms standard sampling methods on a variety of datasets.

A crucial technique for scaling kernel methods to very large data sets reaching or exceeding millions of instances is based on low-rank approximation of kernel matrices. We introduce a new family of algorithms based on mixtures of Nystrom approximations, ensemble Nystrom algorithms, that yield more accurate low-rank approximations than the standard Nystrom method. We give a detailed study of multiple variants of these algorithms based on simple averaging, an exponential weight method, or regression-based methods. We also present a theoretical analysis of these algorithms, including novel error bounds guaranteeing a better convergence rate than the standard Nystrom method. Finally, we report the results of extensive experiments with several data sets containing up to 1M points demonstrating the significant performance improvements gained over the standard Nystrom approximation.

We address the problem of tracking and recognizing faces in real-world, noisy videos. We track faces using a tracker that adaptively builds a target model reflecting changes in appearance, typical of a video setting. However, adaptive appearance trackers often suffer from drift, a gradual adaptation of the tracker to non-targets. To alleviate this problem, our tracker introduces visual constraints using a combination of generative and discriminative models in a particle filtering framework. The generative term conforms the particles to the space of generic face poses while the discriminative one ensures rejection of poorly aligned targets. This leads to a tracker that significantly improves robustness against abrupt appearance changes and occlusions, critical for the subsequent recognition phase. Identity of the tracked subject is established by fusing pose-discriminant and person-discriminant features over the duration of a video sequence. This leads to a robust video-based face recognizer with state-of-the-art recognition performance. We test the quality of tracking and face recognition on real-world noisy videos from YouTube as well as the standard Honda/UCSD database. Our approach produces successful face tracking results on over 80% of all videos without video or person-specific parameter tuning. The good tracking performance induces similarly high recognition rates: 100% on Honda/UCSD and over 70% on the YouTube set containing 35 celebrities in 1500 sequences.

This paper examines the problem of extracting low-dimensional manifold structure given millions of high-dimensional face images. Specifically, we address the computational challenges of nonlinear dimensionality reduction via Isomap and Laplacian Eigenmaps, using a graph containing about 18 million nodes and 65 million edges. Since most manifold learning techniques rely on spectral decomposition, we first analyze two approximate spectral decomposition techniques for large dense matrices (Nystrom and column-sampling), providing the first direct theoretical and empirical comparison between these techniques. We next show extensive experiments on learning low-dimensional embeddings for two large face datasets: CMU-PIE (35 thousand faces) and a web dataset (18 million faces). Our comparisons show that the Nystrom approximation is superior to the column-sampling method. Furthermore, approximate Isomap tends to perform better than Laplacian Eigenmaps on both clustering and classification with the labeled CMU-PIE dataset.

In many vision problems, instead of having fully labeled training data it is easier to obtain the input in small groups, where the data in each group is constrained to be from the same class but the actual class label is not known. Such constraints give rise to partial equivalence relations. The absence of class labels prevents the use of standard discriminative methods in this scenario. On the other hand, the state-of-the-art techniques that use partial equivalence relations, e.g., relevant component analysis, learn projections that are optimal for data representation, but not discrimination. We show that this leads to poor performance in several real-world applications, especially those with high-dimensional data. In this paper, we present a novel discriminative technique for the classification of weakly-labeled data which exploits the null-space of data scatter matrices to achieve good classification accuracy. We demonstrate the superior performance of both linear and nonlinear versions of our approach on face recognition, clustering, and image retrieval tasks. Results are reported on standard datasets as well as real-world images and videos from the Web.

This paper addresses the novel problem of automatically synthesizing an output image from a large collection of different input images. The synthesized image, called a digital tapestry, can be viewed as a visual summary or a virtual 'thumbnail' of all the images in the input collection. The problem of creating the tapestry is cast as a multi-class labeling problem such that each region in the tapestry is constructed from input image blocks that are salient and such that neighboring blocks satisfy spatial compatibility. This is formulated using a Markov random field and optimized via the graph cut based expansion move algorithm. The standard expansion move algorithm can only handle energies with metric terms, while our energy contains non-metric (soft and hard) constraints. Therefore we propose two novel contributions. First, we extend the expansion move algorithm for energy functions with non-metric hard constraints. Secondly, we modify it for functions with "almost" metric soft terms, and show that it gives good results in practice. The proposed framework was tested on several consumer photograph collections, and the results are presented.

We present a two-layer hierarchical formulation to exploit different levels of contextual information in images for robust classification. Each layer is modeled as a conditional field that allows one to capture arbitrary observation-dependent label interactions. The proposed framework has two main advantages. First, it encodes both the short-range interactions (e.g., pixelwise label smoothing) as well as the long-range interactions (e.g., relative configurations of objects or regions) in a tractable manner. Second, the formulation is general enough to be applied to different domains ranging from pixelwise image labeling to contextual object detection. The parameters of the model are learned using a sequential maximum-likelihood approximation. The benefits of the proposed framework are demonstrated on four different datasets and comparison results are presented

We describe an approach that integrates midrange sensing into a dynamic path planning algorithm. The algorithm is based on measuring the reduction in path cost that would be caused by taking a sensor reading from candidate locations. The planner uses this measure in order to decide where to take the next sensor reading. Ideally, one would like to evaluate a path based on a map that is as close as possible to the true underlying world. In practice, however, the map is only sparsely populated by data derived from sensor readings. A key component of the approach described in this paper is a mechanism to infer (or "hallucinate") more complete maps from sparse sensor readings. We show how this hallucination mechanism is integrated with the planner to produce better estimates of the gain in path cost occurred when taking sensor readings. We show results on a real robot as well as a statistical analysis on a large set of randomly generated path planning problems on elevation maps from real terrain.

This paper presents a generative model based approach to man-made structure detection in 2D (two-dimensional) natural images. The proposed approach uses a causal multiscale random field suggested by Bouman and Shapiro (1994) as a prior model on the class labels on the image sites. However, instead of assuming the conditional independence of the observed data, we propose to capture the local dependencies in the data using a multiscale feature vector. The distribution of the multiscale feature vectors is modeled as mixture of Gaussians. A set of robust multi-scale features is presented that captures the general statistical properties of man-made structures at multiple scales without relying on explicit edge detection. The proposed approach was validated on real-world images from the Corel data set, and a performance comparison with other techniques is presented.

In this work we present discriminative random fields (DRFs), a discriminative framework for the classification of image regions by incorporating neighborhood interactions in the labels as well as the observed data. The discriminative random fields offer several advantages over the conventional Markov random field (MRF) framework. First, the DRFs allow to relax the strong assumption of conditional independence of the observed data generally used in the MRF framework for tractability. This assumption is too restrictive for a large number of applications in vision. Second, the DRFs derive their classification power by exploiting the probabilistic discriminative models instead of the generative models used in the MRF framework. Finally, all the parameters in the DRF model are estimated simultaneously from the training data unlike the MRF framework where likelihood parameters are usually learned separately from the field parameters. We illustrate the advantages of the DRFs over the MRF framework in an application of man-made structure detection in natural images taken from the Corel database.

In this paper we present Discriminative Random Fields (DRF), a discrim- inative framework for the classification of natural image regions by incor- porating neighborhood spatial dependencies in the labels as well as the observed data. The proposed model exploits local discriminative models and allows to relax the assumption of conditional independence of the observed data given the labels, commonly used in the Markov Random Field (MRF) framework. The parameters of the DRF model are learned using penalized maximum pseudo-likelihood method. Furthermore, the form of the DRF model allows the MAP inference for binary classifica- tion problems using the graph min-cut algorithms. The performance of the model was verified on the synthetic as well as the real-world images. The DRF model outperforms the MRF model in the experiments.

Images captured with a typical endoscope show spatial distortion, which necessitates distortion correction for subsequent analysis. In this paper, a new methodology based on least squares estimation is proposed to correct the nonlinear distortion in the endoscopic images. A mathematical model based on polynomial mapping is used to map the images from distorted image space onto the corrected image space. The model parameters include the polynomial coefficients, distortion center, and corrected center. The proposed method utilizes a line search approach of global convergence for the iterative procedure to obtain the optimum expansion coefficients. A new technique to find the distortion center of the image based on curvature criterion is presented. A dual-step approach comprising token matching and integrated neighborhood search is also proposed for accurate extraction of the centers of the dots contained in a rectangular grid, used for the model parameter estimation. The model parameters were verified with different grid patterns. The distortion-correction model is applied to several gastrointestinal images and the results are presented. The proposed technique provides high-speed response and forms a key step toward online camera calibration, which is required for accurate quantitative analysis of the images.