Alejandro Hernández-Cano, Alexander Hägele, Allen Hao Huang, Angelika Romanou, Antoni-Joan Solergibert, Barna Pásztor, Bettina Messmer, Dhia Garbaya, Eduard Frank Ďurech, Ido Hakimi, Juan Garcia Giraldo, Mete Ismayilzada, Negar Foroutan, Skander Moalla, Tiancheng Chen, Vinko Sabolčec, Yixuan Xu, Michael Aerni, Badr AlKhamissi, Inés Altemir Marinas, Mohammad Hossein Amani, Matin Ansaripour, Ilia Badanin, Harold Benoit, Emanuela Boros, Nicholas John Browning, Fabian Bösch, Maximilian Böther, Niklas Canova, Camille Challier, Clément Charmillot, Jonathan Coles, Jan Milan Deriu, Arnout Devos, Lukas Drescher, Daniil Dzenhaliou, Maud Ehrmann, Dongyang Fan, Simin Fan, Silin Gao, Miguel Gila, María Grandury, Diba Hashemi, Alexander Miserlis Hoyle, Jiaming Jiang, Mark Klein, Andrei Kucharavy, Anastasiia Kucherenko, Frederike Lübeck, Roman Machacek, Theofilos Ioannis Manitaras, Andreas Marfurt, Kyle Matoba, Simon Matrenok, Henrique Mendonça, Fawzi Roberto Mohamed, Syrielle Montariol, Luca Mouchel, Sven Najem-Meyer, Jingwei Ni, Gennaro Oliva, Matteo Pagliardini, Elia Palme, Andrei Panferov, Léo Paoletti, Marco Passerini, Ivan Pavlov, Auguste Poiroux, Kaustubh Ponkshe, Nathan Ranchin, Javier Rando, Mathieu Sauser, Jakhongir Saydaliev, Mukhammadali Sayfiddinov, Marian Schneider, Stefano Schuppli, Marco Scialanga, Andrei Semenov, Kumar Shridhar, Raghav Singhal, Anna Sotnikova, Alexander Sternfeld, Ayush Kumar Tarun, Paul Teiletche, Jannis Vamvas, Xiaozhe Yao, Hao Zhao, Alexander Ilic, Ana Klimovic, Andreas Krause, Caglar Gulcehre, David Rosenthal, Elliott Ash, Florian Tramèr, Joost VandeVondele, Livio Veraldi, Martin Rajman, Thomas C. Schulthess, Torsten Hoefler, Antoine Bosselut, Martin Jaggi, Imanol Schlag. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

We study stochastic second-order methods for solving general non-convex optimization problems. We propose using a special version of momentum to stabilize the stochastic gradient and Hessian estimates in Newton’s method. We show that momentum provably improves the variance of stochastic estimates and allows the method to converge for any noise level. Using the cubic regularization technique, we prove a global convergence rate for our method on general non-convex problems to a second-order stationary point, even when using only a single stochastic data sample per iteration. This starkly contrasts with all existing stochastic second-order methods for non-convex problems, which typically require large batches. Therefore, we are the first to demonstrate global convergence for batches of arbitrary size in the non-convex case for the Stochastic Cubic Newton. Additionally, we show improved speed on convex stochastic problems for our regularized Newton methods with momentum.

The increasing size of deep neural networks (DNNs) necessitates effective model compression to reduce their computational and memory footprints. Sparsity and quantization are two prominent compression methods that have been shown to reduce DNNs' computational and memory footprints significantly while preserving model accuracy. However, how these two methods interact when combined together remains a key question for developers, as many tacitly assume that they are orthogonal, meaning that their combined use does not introduce additional errors beyond those introduced by each method independently. In this paper, we provide the first mathematical proof that sparsity and quantization are non-orthogonal. We corroborate these results with experiments spanning a range of large language models, including the OPT and LLaMA model families (with 125M to 8B parameters), and vision models like ViT and ResNet. We show that the order in which we apply these methods matters because applying quantization before sparsity may disrupt the relative importance of tensor elements, which may inadvertently remove significant elements from a tensor. More importantly, we show that even if applied in the correct order, the compounded errors from sparsity and quantization can significantly harm accuracy. Our findings extend to the efficient deployment of large models in resource-constrained compute platforms to reduce serving cost, offering insights into best practices for applying these compression methods to maximize hardware resource efficiency without compromising accuracy.

Attention-based transformers have achieved tremendous success across a variety of disciplines including natural languages. To deepen our understanding of their sequential modeling capabilities, there is a growing interest in using Markov input processes to study them. A key finding is that when trained on first-order Markov chains, transformers with two or more layers consistently develop an induction head mechanism to estimate the in-context bigram conditional distribution. In contrast, single-layer transformers, unable to form an induction head, directly learn the Markov kernel but often face a surprising challenge: they become trapped in local minima representing the unigram distribution, whereas deeper models reliably converge to the ground-truth bigram. While single-layer transformers can theoretically model first-order Markov chains, their empirical failure to learn this simple kernel in practice remains a curious phenomenon. To explain this contrasting behavior of single-layer models, in this paper we introduce a new framework for a principled analysis of transformers via Markov chains. Leveraging our framework, we theoretically characterize the loss landscape of single-layer transformers and show the existence of global minima (bigram) and bad local minima (unigram) contingent on data properties and model architecture. We precisely delineate the regimes under which these local optima occur. Backed by experiments, we demonstrate that our theoretical findings are in congruence with the empirical results. Finally, we outline several open problems in this arena.

Scaling language models to larger and deeper sizes has led to significant boosts in performance. Even though the size of these models limits their application in compute-constrained environments, the race to continually develop ever larger and deeper foundational models is underway. At the same time---regardless of the model size---task-specific techniques continue to play a pivotal role in achieving optimal downstream performance. One of these techniques, called Chain-of-Thought (CoT), is particularly interesting since, as we point out in this work, it resembles employing a deeper transformer through re-applying the model multiple times. However, a key subtlety in computing the attention of past tokens differentiates CoT from simply applying the model several times. Based on this insight, we propose CoTFormer, a novel architecture which closely mimics CoT at the token level, allowing us to obtain significantly improved accuracies close to much larger models. While applying CoT introduces additional computation costs, we compensate for it by leveraging CoTFormer's special compatibility with token-wise variable depth. Through a compute adaptive model---which automatically allocates the compute to tokens that need it most---we show that it is possible to reduce the computation cost significantly without any reduction in accuracy, and with further compute cost reductions possible while maintaining a competitive accuracy.

In human-centric settings like education or healthcare, model accuracy and model explainability are key factors for user adoption. Towards these two goals, intrinsically interpretable deep learning models have gained popularity, focusing on accurate predictions alongside faithful explanations. However, there exists a gap in the human-centeredness of these approaches, which often produce nuanced and complex explanations that are not easily actionable for downstream users. We present InterpretCC (interpretable conditional computation), a family of intrinsically interpretable neural networks at a unique point in the design space that optimizes for ease of human understanding and explanation faithfulness, while maintaining comparable performance to state-of-the-art models. InterpretCC achieves this through adaptive sparse activation of features before prediction, allowing the model to use a different, minimal set of features for each instance. We extend this idea into an interpretable, global mixture-of-experts (MoE) model that allows users to specify topics of interest, discretely separates the feature space for each data point into topical subnetworks, and adaptively and sparsely activates these topical subnetworks for prediction. We apply InterpretCC for text, time series and tabular data across several real-world datasets, demonstrating comparable performance with non-interpretable baselines and outperforming intrinsically interpretable baselines. Through a user study involving 56 teachers, InterpretCC explanations are found to have higher actionability and usefulness over other intrinsically interpretable approaches.

On-device LLMs have gained increasing attention for their ability to enhance privacy and provide a personalized user experience. To facilitate private learning with scarce data, Federated Learning has become a standard approach. However, it faces challenges such as computational resource heterogeneity and data heterogeneity among end users. We propose CoMiGS ($\textbf{Co}$llaborative learning with a $\textbf{Mi}$xture of $\textbf{G}$eneralists and $\textbf{S}$pecialists), the first approach to address both challenges. A key innovation of our method is the bi-level optimization formulation of the Mixture-of-Experts learning objective, where the router is optimized using a separate validation set to ensure alignment with the target distribution. We solve our objective with alternating minimization, for which we provide a theoretical analysis. Our method shares generalist experts across users while localizing a varying number of specialist experts, thereby adapting to users’ computational resources and preserving privacy. Through extensive experiments, we show CoMiGS effectively balances general and personalized knowledge for each token generation. We demonstrate that CoMiGS remains robust against overfitting—due to the generalists' regularizing effect—while adapting to local data through specialist expertise. We open source our codebase for collaborative LLMs.

The performance of large language models (LLMs) across diverse downstream applications is fundamentally governed by the quality and composition of their pretraining corpora. Existing domain reweighting algorithms primarily optimize data mixtures for a single target task, thereby resulting in models that overfit to specialized objectives while exhibiting substantial performance degradation on other benchmarks. This paper introduces $\textbf{G}$roup $\textbf{R}$obust Multi-target $\textbf{A}$daptive $\textbf{P}$r$\textbf{E}$training (GRAPE), a novel multi-source-multi-target domain reweighting framework designed to calibrate pretraining data mixtures for robust performance across multiple target tasks simultaneously. GRAPE dynamically adjusts sampling weights across source domains ($\textit{domain weights}$) while concurrently modulating $\textit{task weights}$ that quantify the relative importance of each individual target task. This adaptive process prioritizes tasks based on their learning difficulty throughout training. We formulate this interleaved reweighting mechanism as a minimax optimization problem: The inner maximization adjusts task weights leveraging group distributed-robust-optimization (DRO), where those tasks demonstrating the least improvement under the current data mixture are prioritized with higher weights; The outer minimization then optimizes domain weights to maximize loss reduction on the prioritized tasks. Experiments on $\texttt{ClimbLab}$ and $\texttt{SlimPajama}$ datasets demonstrate that GRAPE consistently outperforms baseline methods in terms of reasoning accuracies across 6 benchmarks. Furthermore, when applied to multilingual targets, GRAPE effectively identifies optimal training mixtures from mainstream languages, achieving superior language modeling capabilities across 8 low-resource target languages.

Large Language Models (LLMs) are commonly pretrained on vast corpora of text without utilizing contextual metadata such as source, quality, or topic, leading to a context-free learning paradigm. While recent studies suggest that adding metadata like URL information as context (i.e., auxiliary inputs not used in the loss calculation) can improve training efficiency and downstream performance, they offer limited understanding of which types of metadata are truly effective and under what conditions. In this work, we conduct a systematic evaluation and find that not all metadata types contribute equally. Only URL context speeds up training, whereas quality scores and topic/format domain information offer no clear benefit. Furthermore, the improved downstream performances of URL conditioning emerge only when longer prompts are used at inference time. In addition, we demonstrate that context-aware pretraining enables more controllable generation than context-free pretraining, in a classifier-free guidance fashion. Although topic and format metadata do not accelerate training, they are effective for steering outputs, offering human-interpretable control over generation.

Despite the significant potential of FP8 data formats for large language model (LLM) pre-training, their adoption has been limited due to challenges in maintaining stability at scale. Existing approaches often rely on suboptimal fine-grained FP8 kernels or fall back to higher-precision matrix multiplications (GEMMs) in sensitive components, such as attention projections, compromising potential throughput gains. We introduce a new class of LLM architectures that, for the first time, support FP8 computation for all GEMMs within transformer blocks during both forward and backward passes. This enables unprecedented throughput gains, particularly at scale, while matching the downstream performance of standard BF16 training. Our architecture design reduces large outlier activations, promoting stable long-term FP8 training. Additionally, we identify key metrics for monitoring low-precision training and predicting potential future divergences.

Dataset curation has become a basis for strong large language model (LLM) performance.While various rule-based filtering heuristics exist for English and multilingual datasets, model-based filtering techniques have primarily focused on English.To address the disparity stemming from limited research on non-English languages, we develop a model-based filtering framework for multilingual datasets that aims to identify a diverse set of structured and knowledge-rich samples.Our approach emphasizes transparency, simplicity, and efficiency, leveraging Transformer- and FastText-based classifiers to ensure the broad accessibility of our technique and data.We conduct comprehensive ablation studies on the FineWeb-2 web crawl dataset across diverse language families, scripts, and resource availability to demonstrate the effectiveness of our method.Training a 1B-parameter Llama model for 70B and 119B tokens, our approach can match the baseline MMLU score with as little as 15\% of the training tokens, while also improving across other benchmarks and mitigating the curse of multilinguality.These findings provide strong evidence for the generalizability of our approach to other languages. As a result, we extend our framework to 20 languages for which we release the refined pretraining datasets.

Batch Normalization (BN) is widely used to stabilize the optimization process and improve the test performance of deep neural networks. The regularization effect of BN depends on the batch size and explicitly using smaller batch sizes with Batch Normalization, a method known as Ghost Batch Normalization (GBN), has been found to improve generalization in many settings. We investigate the effectiveness of GBN by disentangling the induced ``Ghost Noise'' from normalization and quantitatively analyzing the distribution of noise as well as its impact on model performance. Inspired by our analysis, we propose a new regularization technique called Ghost Noise Injection (GNI) that imitates the noise in GBN without incurring the detrimental train-test discrepancy effects of small batch training. We experimentally show that GNI can provide a greater generalization benefit than GBN. Ghost Noise Injection can also be beneficial in otherwise non-noisy settings such as layer-normalized networks, providing additional evidence of the usefulness of Ghost Noise in Batch Normalization as a regularizer.

Averaging neural network parameters is an intuitive method for fusing the knowledge of two independent models. It is most prominently used in federated learning. If models are averaged at the end of training, this can only lead to a good performing model if the loss surface of interest is very particular, i.e., the loss in the midpoint between the two models needs to be sufficiently low. This is impossible to guarantee for the non-convex losses of state-of-the-art networks. For averaging models trained on vastly different datasets, it was proposed to average only the parameters of particular layers or combinations of layers, resulting in better performing models. To get a better understanding of the effect of layer-wise averaging, we analyse the performance of the models that result from averaging single layers, or groups of layers. Based on our empirical and theoretical investigation, we introduce a novel notion of the layer-wise linear connectivity, and show that deep networks do not have layer-wise barriers between them.

Decentralized learning is appealing as it enables the scalable usage of large amounts of distributed data and resources without resorting to any central entity, while promoting privacy since every user minimizes the direct exposure of their data. Yet, without additional precautions, curious users can still leverage models obtained from their peers to violate privacy. In this paper, we propose Decor, a variant of decentralized SGD with differential privacy (DP) guarantees. Essentially, in Decor, users securely exchange randomness seeds in one communication round to generate pairwise-canceling correlated Gaussian noises, which are injected to protect local models at every communication round. We theoretically and empirically show that, for arbitrary connected graphs, Decor matches the central DP optimal privacy-utility trade-off. We do so under SecLDP, our new relaxation of local DP, which protects all user communications against an external eavesdropper and curious users, assuming that every pair of connected users shares a secret, i.e., an information hidden to all others. The main theoretical challenge is to control the accumulation of non-canceling correlated noise due to network sparsity. We also propose a companion SecLDP privacy accountant for public use.

The performance of optimization methods is often tied to the spectrum of the objective Hessian. Yet, conventional assumptions, such as smoothness, do often not enable us to make finely-grained convergence statements—particularly not for non-convex problems. Striving for a more intricate characterization of complexity, we introduce a unique concept termed graded non-convexity. This allows to partition the class of non-convex problems into a nested chain of subclasses. Interestingly, many traditional non-convex objectives, including partially convex problems, matrix factorizations, and neural networks, fall within these subclasses. As a second contribution, we propose gradient methods with spectral preconditioning, which employ inexact top eigenvectors of the Hessian to address the ill-conditioning of the problem, contingent on the grade. Our analysis reveals that these new methods provide provably superior convergence rates compared to basic gradient descent on applicable problem classes, particularly when large gaps exist between the top eigenvalues of the Hessian. Our theory is validated by numerical experiments executed on multiple practical machine learning problems.

The coverage and composition of the pretraining data significantly impacts the generalization ability of Large Language Models (LLMs). Despite its importance, recent LLMs still rely on heuristics and trial and error to increase or reduce the influence of data-domains. We propose DOmain reweighting with Generalization Estimation (DoGE), which optimizes the probability of sampling from each domain (domain weights) in a principled way. Our approach is a two stage process consisting (i) training a proxy model to obtain domain weights using a bi-level optimization algorithm; (ii) training a larger base model by sampling training domains according to the learnt domain weights. In our experiments, we extensively show how DoGE improves the generalization of the base model to any target data mixture. On the SlimPajama dataset, our base model gets a better perplexity and few-shot reasoning accuracies across 6 tasks compared to baseline methods. Moreover, aiming to generalize to out-of-domain target tasks, which is unseen in the pretraining corpus (OOD domain), DoGE can effectively identify inter-domain dependencies, consistently achieves better test perplexity on the target domain.

In machine learning and neural network optimization, algorithms like incremental gradient, single shuffle SGD, and random reshuffle SGD are popular due to their cache-mismatch efficiency and good practical convergence behavior. However, their optimization properties in theory, especially for non-convex smooth functions, remain incompletely explored. This paper delves into the convergence properties of SGD algorithms with arbitrary data ordering, within a broad framework for non-convex smooth functions. Our findings show enhanced convergence guarantees for incremental gradient and single shuffle SGD. Particularly if $n$ is the training set size, we improve $n$ times the optimization term of convergence guarantee to reach accuracy $\epsilon$ from $O \left( \frac{n}{\epsilon} \right)$ to $O \left( \frac{1}{\epsilon}\right)$.

This study investigates how weight decay affects the update behavior of individual neurons in deep neural networks through a combination of applied analysis and experimentation. Weight decay can cause the expected magnitude and angular updates of a neuron’s weight vector to converge to a steady state we call rotational equilibrium. These states can be highly homogeneous, effectively balancing the average rotation—a proxy for the effective learning rate—across different layers and neurons. Our work analyzes these dynamics across optimizers like Adam, Lion, and SGD with momentum, offering a new simple perspective on training that elucidates the efficacy of widely used but poorly understood methods in deep learning. We demonstrate how balanced rotation plays a key role in the effectiveness of normalization like Weight Standardization, as well as that of AdamW over Adam with L2-regularization. Finally, we show that explicitly controlling the rotation provides the benefits of weight decay while substantially reducing the need for learning rate warmup.

Data-parallel SGD is the de facto algorithm for distributed optimization, especially for large scale machine learning. Despite its merits, communication bottleneck is one of its persistent issues. Most compression schemes to alleviate this either assume noiseless communication links, or fail to achieve good performance on practical tasks. In this paper, we close this gap and introduce **LASER**: **L**ine**A**r Compre**S**sion in Wir**E**less Dist**R**ibuted Optimization. LASER capitalizes on the inherent low-rank structure of gradients and transmits them efficiently over the noisy channels. Whilst enjoying theoretical guarantees similar to those of the classical SGD, LASER shows consistent gains over baselines on a variety of practical benchmarks. In particular, it outperforms the state-of-the-art compression schemes on challenging computer vision and GPT language modeling tasks. On the latter, we obtain 50-64% improvement in perplexity over our baselines for noisy channels.

We introduce QuaRot, a new Quantization scheme based on Rotations, which is able to quantize LLMs end-to-end, including all weights, activations, and KV cache in 4 bits. QuaRot rotates LLMs in a way that removes outliers from the hidden state without changing the output, making quantization easier. This computational invariance is applied to the hidden state (residual) of the LLM, as well as to the activations of the feed-forward components, aspects of the attention mechanism, and to the KV cache. The result is a quantized model where all matrix multiplications are performed in 4 bits, without any channels identified for retention in higher precision. Our 4-bit quantized LLAMA2-70B model has losses of at most 0.47 WikiText-2 perplexity and retains 99% of the zero-shot performance. We also show that QuaRot can provide lossless 6 and 8 bit LLAMA-2 models without any calibration data using round-to-nearest quantization. Code is available at github.com/spcl/QuaRot.

Scale has become a main ingredient in obtaining strong machine learning models. As a result, understanding a model's scaling properties is key to effectively designing both the right training setup as well as future generations of architectures. In this work, we argue that scale and training research has been needlessly complex due to reliance on the cosine schedule, which prevents training across different lengths for the same model size. We investigate the training behavior of a direct alternative --- constant learning rate and cooldowns --- and find that it scales predictably and reliably similar to cosine. Additionally, we show that stochastic weight averaging yields improved performance along the training trajectory, without additional training costs, across different scales. Importantly, with these findings we demonstrate that scaling experiments can be performed with significantly reduced compute and GPU hours by utilizing fewer but reusable training runs. Our code is available at https://github.com/epfml/schedules-and-scaling/.

Collaborative learning is an important tool to train multiple clients more effectively by enabling communication among clients. Identifying helpful clients, however, presents challenging and often introduces significant overhead. In this paper, we model **client-selection** and **model-training** as two interconnected optimization problems, proposing a novel bilevel optimization problem for collaborative learning. We introduce **CoBo**, a *scalable* and *elastic*, SGD-type alternating optimization algorithm that efficiently addresses these problem with theoretical convergence guarantees. Empirically, **CoBo** achieves superior performance, surpassing popular personalization algorithms by 9.3% in accuracy on a task with high heterogeneity, involving datasets distributed among 80 clients.

Learning Rate Warmup is a popular heuristic for training neural networks, especially at larger batch sizes, despite limited understanding of its benefits. Warmup decreases the update size $\Delta \mathbf{w}_t = \eta_t \mathbf{u}_t$ early in training by using lower values for the learning rate $\eta_t$. In this work we argue that warmup benefits training by keeping the overall size of $\Delta \mathbf{w}_t$ limited, counteracting large initial values of $\mathbf{u}_t$. Focusing on small-scale GPT training with AdamW/Lion, we explore the following question: *Why and by which criteria are early updates $\mathbf{u}_t$ too large?* We analyze different metrics for the update size including the $\ell_2$-norm, resulting directional change, and impact on the representations of the network, providing a new perspective on warmup. In particular, we find that warmup helps counteract large angular updates as well as a limited critical batch size early in training. Finally, we show that the need for warmup can be significantly reduced or eliminated by modifying the optimizer to explicitly normalize $\mathbf{u}_t$ based on the aforementioned metrics.

The transformer architecture by Vaswani et al. (2017) is now ubiquitous across application domains, from natural language processing to speech processing and image understanding. We propose DenseFormer, a simple modification to the standard architecture that improves the perplexity of the model without increasing its size---adding a few thousand parameters for large-scale models in the 100B parameters range. Our approach relies on an additional averaging step after each transformer block, which computes a weighted average of current and past representations---we refer to this operation as Depth-Weighted-Average (DWA). The learned DWA weights exhibit coherent patterns of information flow, revealing the strong and structured reuse of activations from distant layers. Experiments demonstrate that DenseFormer is more data efficient, reaching the same perplexity of much deeper transformer models, and that for the same perplexity, these new models outperform transformer baselines in terms of memory efficiency and inference time.

Document-level text simplification is a specific type of simplification which involves simplifying documents consisting of several sentences by rewriting them into fewer or more sentences.In this paper, we propose a new two-stage framework SIMSUM for automated document-level text simplification.Our model is designed with explicit summarization and simplification models and guides the generation using the main keywords of a source text.In order to evaluate our new model, we use two existing benchmark datasets for simplification, namely D-Wikipedia and Wiki-Doc.We compare our model's performance with state of the art and show that SIMSUM achieves top results on the D-Wikipedia dataset SARI (+1.20),D-SARI (+1.64), and FKGL (-0.35) scores, improving over the best baseline models.In order to evaluate the quality of the generated text, we analyze the outputs from different models qualitatively and demonstrate the merit of our new model.Our code and datasets are available 1 .

This paper studies first-order algorithms for solving fully composite optimization problems over convex and compact sets. We leverage the structure of the objective by handling its differentiable and non-differentiable components separately, linearizing only the smooth parts. This provides us with new generalizations of the classical Frank-Wolfe method and the Conditional Gradient Sliding algorithm, that cater to a subclass of non-differentiable problems. Our algorithms rely on a stronger version of the linear minimization oracle, which can be efficiently implemented in several practical applications. We provide the basic version of our method with an affine-invariant analysis and prove global convergence rates for both convex and non-convex objectives. Furthermore, in the convex case, we propose an accelerated method with correspondingly improved complexity. Finally, we provide illustrative experiments to support our theoretical results.

We analyze Newton’s method with lazy Hessian updates for solving general possibly non-convex optimization problems. We propose to reuse a previously seen Hessian for several iterations while computing new gradients at each step of the method. This significantly reduces the overall arithmetic complexity of second-order optimization schemes. By using the cubic regularization technique, we establish fast global convergence of our method to a second-order stationary point, while the Hessian does not need to be updated each iteration. For convex problems, we justify global and local superlinear rates for lazy Newton steps with quadratic regularization, which is easier to compute. The optimal frequency for updating the Hessian is once every $d$ iterations, where $d$ is the dimension of the problem. This provably improves the total arithmetic complexity of second-order algorithms by a factor $\sqrt{d}$.

When training neural networks, it has been widely observed that a large step size is essential in stochastic gradient descent (SGD) for obtaining superior models. However, the effect of large step sizes on the success of SGD is not well understood theoretically. Several previous works have attributed this success to the stochastic noise present in SGD. However, we show through a novel set of experiments that the stochastic noise is not sufficient to explain good non-convex training, and that instead the effect of a large learning rate itself is essential for obtaining best performance.We demonstrate the same effects also in the noise-less case, i.e. for full-batch GD. We formally prove that GD with large step size —on certain non-convex function classes — follows a different trajectory than GD with a small step size, which can lead to convergence to a global minimum instead of a local one. Our settings provide a framework for future analysis which allows comparing algorithms based on behaviors that can not be observed in the traditional settings.

We consider a collaborative learning setting where the goal of each agent is to improve their own model by leveraging the expertise of collaborators, in addition to their own training data. To facilitate the exchange of expertise among agents, we propose a distillation-based method leveraging shared unlabeled auxiliary data, which is pseudo-labeled by the collective. Central to our method is a trust weighting scheme that serves to adaptively weigh the influence of each collaborator on the pseudo-labels until a consensus on how to label the auxiliary data is reached. We demonstrate empirically that our collaboration scheme is able to significantly boost individual models’ performance in the target domain from which the auxiliary data is sampled. At the same time, it can provably mitigate the negative impact of bad models on the collective. By design, our method adeptly accommodates heterogeneity in model architectures and substantially reduces communication overhead compared to typical collaborative learning methods.

Multiplications are responsible for most of the computational cost involved in neural network training and inference. Recent research has thus looked for ways to reduce the cost associated with them. Inspired by Mogami 2020, we replace multiplication with a cheap piecewise affine approximation that is achieved by adding the bit representation of the floating point numbers together as integers. We show that transformers can be trained with the resulting modified matrix multiplications on both vision and language tasks with little to no performance impact, and without changes to the training hyperparameters. We further replace all non-linearities in the networks making them fully and jointly piecewise affine in both inputs and weights. Finally, we show that we can eliminate all multiplications in the entire training process, including operations in the forward pass, backward pass and optimizer update, demonstrating the first successful training of modern neural network architectures in a fully multiplication-free fashion.

While Transformers have shown remarkable success in natural language processing, their attention mechanism's large memory requirements have limited their ability to handle longer contexts. Prior approaches, such as recurrent memory or retrieval-based augmentation, have either compromised the random-access flexibility of attention (i.e., the capability to select any token in the entire context) or relied on separate mechanisms for relevant context retrieval, which may not be compatible with the model's attention. In this paper, we present a novel approach that allows access to the complete context while retaining random-access flexibility, closely resembling running attention on the entire context. Our method uses a landmark token to represent each block of the input and trains the attention to use it for selecting relevant blocks, enabling retrieval of blocks directly through the attention mechanism instead of by relying on a separate mechanism. Our approach seamlessly integrates with specialized data structures and the system's memory hierarchy, enabling processing of arbitrarily long context lengths. We demonstrate that our method can obtain comparable performance with Transformer-XL while significantly reducing the number of retrieved tokens in each step. Finally, we show that fine-tuning LLaMA 7B with our method successfully extends its context length capacity to over 32k tokens, allowing for inference at the context lengths of GPT-4. We release the implementation of landmark attention and the code to reproduce our experiments at https://github.com/epfml/landmark-attention/.

Transformer-based language models have found many diverse applications requiring them to process sequences of increasing length. For these applications, the causal self-attention---which is the only component scaling quadratically w.r.t. the sequence length---becomes a central concern. While many works have proposed schemes to sparsify the attention patterns and reduce the computational overhead of self-attention, those are often limited by implementation concerns and end up imposing a simple and static structure over the attention matrix. Conversely, implementing more dynamic sparse attention often results in runtimes significantly slower than computing the full attention using the Flash implementation from Dao et al. (2022). We extend FlashAttention to accommodate a large class of attention sparsity patterns that, in particular, encompass key/query dropping and hashing-based attention. This leads to implementations with no computational complexity overhead and a multi-fold runtime speedup on top of FlashAttention. Even with relatively low degrees of sparsity, our method improves visibly upon FlashAttention as the sequence length increases. Without sacrificing perplexity, we increase the training speed of a transformer language model by $2.0\times$ and $3.3\times$ for sequences of respectively $8k$ and $16k$ tokens.

Predicting multiple real-world tasks in a single model often requires a particularly diverse feature space. Multimodal (MM) models aim to extract the synergistic predictive potential of multiple data types to create a shared feature space with aligned semantic meaning across inputs of drastically varying sizes (i.e. images, text, sound). Most current MM architectures fuse these representations in parallel, which not only limits their interpretability but also creates a dependency on modality availability. We present MultiModN, a multimodal, modular network that fuses latent representations in a sequence of any number, combination, or type of modality while providing granular real-time predictive feedback on any number or combination of predictive tasks. MultiModN's composable pipeline is interpretable-by-design, as well as innately multi-task and robust to the fundamental issue of biased missingness. We perform four experiments on several benchmark MM datasets across 10 real-world tasks (predicting medical diagnoses, academic performance, and weather), and show that MultiModN's sequential MM fusion does not compromise performance compared with a baseline of parallel fusion. By simulating the challenging bias of missing not-at-random (MNAR), this work shows that, contrary to MultiModN, parallel fusion baselines erroneously learn MNAR and suffer catastrophic failure when faced with different patterns of MNAR at inference. To the best of our knowledge, this is the first inherently MNAR-resistant approach to MM modeling. In conclusion, MultiModN provides granular insights, robustness, and flexibility without compromising performance.

In data-parallel optimization of machine learning models, workers collaborate to improve their estimates of the model: more accurate gradients allow them to use larger learning rates and optimize faster. In the decentralized setting, in which workers communicate over a sparse graph, current theory fails to capture important aspects of real-world behavior. First, the `spectral gap' of the communication graph is not predictive of its empirical performance in (deep) learning. Second, current theory does not explain that collaboration enables larger learning rates than training alone. In fact, it prescribes smaller learning rates, which further decrease as graphs become larger, failing to explain convergence dynamics in infinite graphs. This paper aims to paint an accurate picture of sparsely-connected distributed optimization. We quantify how the graph topology influences convergence in a quadratic toy problem and provide theoretical results for general smooth and (strongly) convex objectives. Our theory matches empirical observations in deep learning, and accurately describes the relative merits of different graph topologies.

A major obstacle to achieving global convergence in distributed and federated learning is the misalignment of gradients across clients or mini-batches due to heterogeneity and stochasticity of the distributed data. In this work, we show that data heterogeneity can in fact be exploited to improve generalization performance through implicit regularization. One way to alleviate the effects of heterogeneity is to encourage the alignment of gradients across different clients throughout training. Our analysis reveals that this goal can be accomplished by utilizing the right optimization method that replicates the implicit regularization effect of SGD, leading to gradient alignment as well as improvements in test accuracies. Since the existence of this regularization in SGD completely relies on the sequential use of different mini-batches during training, it is inherently absent when training with large mini-batches. To obtain the generalization benefits of this regularization while increasing parallelism, we propose a novel GradAlign algorithm that induces the same implicit regularization while allowing the use of arbitrarily large batches in each update. We experimentally validate the benefits of our algorithm in different distributed and federated learning settings.

State-of-the-art training algorithms for deep learning models are based on stochastic gradient descent (SGD). Recently, many variations have been explored: perturbing parameters for better accuracy (such as in Extragradient), limiting SGD updates to a subset of parameters for increased efficiency (such as meProp) or a combination of both (such as Dropout). However, the convergence of these methods is often not studied in theory. We propose a unified theoretical framework to study such SGD variants—encompassing the aforementioned algorithms and additionally a broad variety of methods used for communication efficient training or model compression. Our insights can be used as a guide to improve the efficiency of such methods and facilitate generalization to new applications. As an example, we tackle the task of jointly training networks, a version of which (limited to sub-networks) is used to create Slimmable Networks. By training a low-rank Transformer jointly with a standard one we obtain superior performance than when it is trained separately.

Fedor Moiseev, Zhe Dong, Enrique Alfonseca, Martin Jaggi. Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies. 2022.

We study the asynchronous stochastic gradient descent algorithm, for distributed training over $n$ workers that might be heterogeneous. In this algorithm, workers compute stochastic gradients in parallel at their own pace and return them to the server without any synchronization.Existing convergence rates of this algorithm for non-convex smooth objectives depend on the maximum delay $\tau_{\max}$ and reach an $\epsilon$-stationary point after $O\!\left(\sigma^2\epsilon^{-2}+ \tau_{\max}\epsilon^{-1}\right)$ iterations, where $\sigma$ is the variance of stochastic gradients. In this work (i) we obtain a tighter convergence rate of $O\!\left(\sigma^2\epsilon^{-2}+ \sqrt{\tau_{\max}\tau_{avg}}\epsilon^{-1}\right)$ *without any change in the algorithm* where $\tau_{avg}$ is the average delay, which can be significantly smaller than $\tau_{\max}$. We also provide (ii) a simple delay-adaptive learning rate scheme, under which asynchronous SGD achieves a convergence rate of $O\!\left(\sigma^2\epsilon^{-2}+ \tau_{avg}\epsilon^{-1}\right)$, and does not require any extra hyperparameter tuning nor extra communications. Our result allows to show *for the first time* that asynchronous SGD is *always faster* than mini-batch SGD. In addition, (iii) we consider the case of heterogeneous functions motivated by federated learning applications and improve the convergence rate by proving a weaker dependence on the maximum delay compared to prior works.

Federated Learning (FL) is a novel approach enabling several clients holding sensitive data to collaboratively train machine learning models, without centralizing data. The cross-silo FL setting corresponds to the case of few ($2$--$50$) reliable clients, each holding medium to large datasets, and is typically found in applications such as healthcare, finance, or industry. While previous works have proposed representative datasets for cross-device FL, few realistic healthcare cross-silo FL datasets exist, thereby slowing algorithmic research in this critical application. In this work, we propose a novel cross-silo dataset suite focused on healthcare, FLamby (Federated Learning AMple Benchmark of Your cross-silo strategies), to bridge the gap between theory and practice of cross-silo FL.FLamby encompasses 7 healthcare datasets with natural splits, covering multiple tasks, modalities, and data volumes, each accompanied with baseline training code. As an illustration, we additionally benchmark standard FL algorithms on all datasets.Our flexible and modular suite allows researchers to easily download datasets, reproduce results and re-use the different components for their research. FLamby is available at~\url{www.github.com/owkin/flamby}.

In data-parallel optimization of machine learning models, workers collaborate to improve their estimates of the model: more accurate gradients allow them to use larger learning rates and optimize faster. We consider the setting in which all workers sample from the same dataset, and communicate over a sparse graph (decentralized). In this setting, current theory fails to capture important aspects of real-world behavior. First, the ‘spectral gap’ of the communication graph is not predictive of its empirical performance in (deep) learning. Second, current theory does not explain that collaboration enables larger learning rates than training alone. In fact, it prescribes smaller learning rates, which further decrease as graphs become larger, failing to explain convergence in infinite graphs. This paper aims to paint an accurate picture of sparsely-connected distributed optimization when workers share the same data distribution. We quantify how the graph topology influences convergence in a quadratic toy problem and provide theoretical results for general smooth and (strongly) convex objectives. Our theory matches empirical observations in deep learning, and accurately describes the relative merits of different graph topologies.

Prakhar Gupta, Martin Jaggi. Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers). 2021.

Zhuoyuan Mao, Prakhar Gupta, Chenhui Chu, Martin Jaggi, Sadao Kurohashi. Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers). 2021.

In distributed optimization, parameter updates from the gradient computing node devices have to be aggregated in every iteration on the orchestrating server. When these updates are sent over an arbitrary commodity network, bandwidth and latency can be limiting factors. We propose a communication framework where nodes may skip unnecessary uploads. Every node locally accumulates an error vector in memory and self-triggers the upload of the memory contents to the parameter server using a significance filter. The server then uses a history of the nodes’ gradients to update the parameter. We characterize the convergence rate of our algorithm in smooth settings (strongly-convex, convex, and non-convex) and show that it enjoys the same convergence rate as when sending gradients every iteration, with substantially fewer uploads. Numerical experiments on real data indicate a significant reduction of used network resources (total communicated bits and latency), especially in large networks, compared to state-of-the-art algorithms. Our results provide important practical insights for using machine learning over resource-constrained networks, including Internet-of-Things and geo-separated datasets across the globe.

Decentralized optimization methods enable on-device training of machine learning models without a central coordinator. In many scenarios communication between devices is energy demanding and time consuming and forms the bottleneck of the entire system. We propose a new randomized first-order method which tackles the communication bottleneck by applying randomized compression operators to the communicated messages. By combining our scheme with a new variance reduction technique that progressively throughout the iterations reduces the adverse effect of the injected quantization noise, we obtain a scheme that converges linearly on strongly convex decentralized problems while using compressed communication only. We prove that our method can solve the problems without any increase in the number of communications compared to the baseline which does not perform any communication compression while still allowing for a significant compression factor which depends on the conditioning of the problem and the topology of the network. We confirm our theoretical findings in numerical experiments.

It has been experimentally observed that the efficiency of distributed training with stochastic gradient (SGD) depends decisively on the batch size and—in asynchronous implementations—on the gradient staleness. Especially, it has been observed that the speedup saturates beyond a certain batch size and/or when the delays grow too large. We identify a data-dependent parameter that explains the speedup saturation in both these settings. Our comprehensive theoretical analysis, for strongly convex, convex and non-convex settings, unifies and generalized prior work directions that often focused on only one of these two aspects. In particular, our approach allows us to derive improved speedup results under frequently considered sparsity assumptions. Our insights give rise to theoretically based guidelines on how the learning rates can be adjusted in practice. We show that our results are tight and illustrate key findings in numerical experiments.

Word embeddings have gained increasing popularity in the recent years due to the Word2vec library and its extension fastText that uses subword information. In this paper, we aim at improving the execution speed of fastText training on homo-geneous multi- and manycore CPUs while maintaining accuracy. We present a novel open-source implementation that flexibly incorporates various algorithmic variants including negative sample sharing, batched updates, and a byte-pair encoding-based alternative for subword units. We build these novel variants over a fastText implementation that we carefully optimized for the architecture, memory hierarchy, and parallelism of current manycore CPUs. Our experiments on three languages demon-strate 3–20 x speed-up in training time at competitive semantic and syntactic accuracy.

Data augmentation is a widely adopted technique for avoiding overfitting when training deep neural networks. However, this approach requires domain-specific knowledge and is often limited to a fixed set of hard-coded transformations. Recently, several works proposed to use generative models for generating semantically meaningful perturbations to train a classifier. However, because accurate encoding and decoding are critical, these methods, which use architectures that approximate the latent-variable inference, remained limited to pilot studies on small datasets.Exploiting the exactly reversible encoder-decoder structure of normalizing flows, we perform on-manifold perturbations in the latent space to define fully unsupervised data augmentations. We demonstrate that such perturbations match the performance of advanced data augmentation techniques—reaching 96.6% test accuracy for CIFAR10 using ResNet-18 and outperform existing methods, particularly in low data regimes—yielding 10–25% relative improvement of test accuracy from classical training. We find that our latent adversarial perturbations adaptive to the classifier throughout its training are most effective, yielding the first test accuracy improvement results on real-world datasets—CIFAR-10/100—via latent-space perturbations.

Decentralized training of deep learning models enables on-device learning over networks, as well as efficient scaling to large compute clusters. Experiments in earlier works reveal that, even in a data-center setup, decentralized training often suffers from the degradation in the quality of the model: the training and test performance of models trained in a decentralized fashion is in general worse than that of models trained in a centralized fashion, and this performance drop is impacted by parameters such as network size, communication topology and data partitioning. We identify the changing consensus distance between devices as a key parameter to explain the gap between centralized and decentralized training. We show in theory that when the training consensus distance is lower than a critical quantity, decentralized training converges as fast as the centralized counterpart. We empirically validate that the relation between generalization performance and consensus distance is consistent with this theoretical observation. Our empirical insights allow the principled design of better decentralized training schemes that mitigate the performance drop. To this end, we provide practical training guidelines and exemplify its effectiveness on the data-center setup as the important first step.

Decentralized training of deep learning models is a key element for enabling data privacy and on-device learning over networks. In realistic learning scenarios, the presence of heterogeneity across different clients’ local datasets poses an optimization challenge and may severely deteriorate the generalization performance. In this paper, we investigate and identify the limitation of several decentralized optimization algorithms for different degrees of data heterogeneity. We propose a novel momentum-based method to mitigate this decentralized training difficulty. We show in extensive empirical experiments on various CV/NLP datasets (CIFAR-10, ImageNet, and AG News) and several network topologies (Ring and Social Network) that our method is much more robust to the heterogeneity of clients’ data than other existing methods, by a significant improvement in test performance (1%-20%).

Conformal Predictors (CP) are wrappers around ML models, providing error guarantees under weak assumptions on the data distribution. They are suitable for a wide range of problems, from classification and regression to anomaly detection. Unfortunately, their very high computational complexity limits their applicability to large datasets. In this work, we show that it is possible to speed up a CP classifier considerably, by studying it in conjunction with the underlying ML method, and by exploiting incremental&decremental learning. For methods such as k-NN, KDE, and kernel LS-SVM, our approach reduces the running time by one order of magnitude, whilst producing exact solutions. With similar ideas, we also achieve a linear speed up for the harder case of bootstrapping. Finally, we extend these techniques to improve upon an optimization of k-NN CP for regression. We evaluate our findings empirically, and discuss when methods are suitable for CP optimization.

Byzantine robustness has received significant attention recently given its importance for distributed and federated learning. In spite of this, we identify severe flaws in existing algorithms even when the data across the participants is identically distributed. First, we show realistic examples where current state of the art robust aggregation rules fail to converge even in the absence of any Byzantine attackers. Secondly, we prove that even if the aggregation rules may succeed in limiting the influence of the attackers in a single round, the attackers can couple their attacks across time eventually leading to divergence. To address these issues, we present two surprisingly simple strategies: a new robust iterative clipping procedure, and incorporating worker momentum to overcome time-coupled attacks. This is the first provably robust method for the standard stochastic optimization setting.

DNN inference accelerators executing online services exhibit low average loads because of service demand variability, leading to poor resource utilization. Unfortunately, reclaiming idle inference cycles is difficult as other workloads can not execute on a custom accelerator. With recent proposals for the use of fixed-point arithmetic in training, there are opportunities for training services to piggyback on inference accelerators. We make the observation that a key challenge in doing so is maintaining service-level latency constraints for inference. We show that relaxing latency constraints in an inference accelerator with ALU arrays that are batching-optimized achieves near-optimal throughput for a given area and power envelope while maintaining inference services’ tail latency goals.

Federated learning (FL) is a challenging setting for optimization due to the heterogeneity of the data across different clients which gives rise to the client drift phenomenon. In fact, obtaining an algorithm for FL which is uniformly better than simple centralized training has been a major open problem thus far. In this work, we propose a general algorithmic framework, Mime, which i) mitigates client drift and ii) adapts arbitrary centralized optimization algorithms such as momentum and Adam to the cross-device federated learning setting. Mime uses a combination of control-variates and server-level statistics (e.g. momentum) at every client-update step to ensure that each local update mimics that of the centralized method run on iid data. We prove a reduction result showing that Mime can translate the convergence of a generic algorithm in the centralized setting into convergence in the federated setting. Further, we show that when combined with momentum based variance reduction, Mime is provably faster than any centralized method--the first such result. We also perform a thorough experimental exploration of Mime's performance on real world datasets.

In decentralized machine learning, workers compute model updates on their local data.Because the workers only communicate with few neighbors without central coordination, these updates propagate progressively over the network.This paradigm enables distributed training on networks without all-to-all connectivity, helping to protect data privacy as well as to reduce the communication cost of distributed training in data centers.A key challenge, primarily in decentralized deep learning, remains the handling of differences between the workers' local data distributions.To tackle this challenge, we introduce the RelaySum mechanism for information propagation in decentralized learning.RelaySum uses spanning trees to distribute information exactly uniformly across all workers with finite delays depending on the distance between nodes.In contrast, the typical gossip averaging mechanism only distributes data uniformly asymptotically while using the same communication volume per step as RelaySum.We prove that RelaySGD, based on this mechanism, is independent of data heterogeneity and scales to many workers, enabling highly accurate decentralized deep learning on heterogeneous data.

We present a framework for building unsupervised representations of entities and their compositions, where each entity is viewed as a probability distribution rather than a vector embedding. In particular, this distribution is supported over the contexts which co-occur with the entity and are embedded in a suitable low-dimensional space. This enables us to consider representation learning from the perspective of Optimal Transport and take advantage of its tools such as Wasserstein distance and barycenters. We elaborate how the method can be applied for obtaining unsupervised representations of text and illustrate the performance (quantitatively as well as qualitatively) on tasks such as measuring sentence similarity, word entailment and similarity, where we empirically observe significant gains (e.g., 4.1% relative improvement over Sent2vec, GenSen).The key benefits of the proposed approach include: (a) capturing uncertainty and polysemy via modeling the entities as distributions, (b) utilizing the underlying geometry of the particular task (with the ground cost), (c) simultaneously providing interpretability with the notion of optimal transport between contexts and (d) easy applicability on top of existing point embedding methods. The code, as well as pre-built histograms, are available under https://github.com/context-mover/.

We present an efficient method of utilizing pretrained language models, where we learn selective binary masks for pretrained weights in lieu of modifying them through finetuning.Extensive evaluations of masking BERT, RoBERTa, and DistilBERT on eleven diverse NLP tasks show that our masking scheme yields performance comparable to finetuning, yet has a much smaller memory footprint when several tasks need to be inferred.Intrinsic evaluations show that representations computed by our binary masked language models encode information necessary for solving downstream tasks.Analyzing the loss landscape, we show that masking and finetuning produce models that reside in minima that can be connected by a line segment with nearly constant test accuracy.This confirms that masking can be utilized as an efficient alternative to finetuning.

Decentralized stochastic optimization methods have gained a lot of attention recently, mainly because of their cheap per iteration cost, data locality, and their communication-efficiency. In this paper we introduce a unified convergence analysis that covers a large variety of decentralized SGD methods which so far have required different intuitions, have different applications, and which have been developed separately in various communities. Our algorithmic framework covers local SGD updates and synchronous and pairwise gossip updates on adaptive network topology. We derive universal convergence rates for smooth (convex and non-convex) problems and the rates interpolate between the heterogeneous (non-identically distributed data) and iid-data settings, recovering linear convergence rates in many special cases, for instance for over-parametrized models. Our proofs rely on weak assumptions (typically improving over prior work in several aspects) and recover (and improve) the best known complexity results for a host of important scenarios, such as for instance coorperative SGD and federated averaging (local SGD).

Deep learning networks are typically trained by Stochastic Gradient Descent (SGD) methods that iteratively improve the model parameters by estimating a gradient on a very small fraction of the training data. A major roadblock faced when increasing the batch size to a substantial fraction of the training data for reducing training time is the persistent degradation in performance (generalization gap). To address this issue, recent work propose to add small perturbations to the model parameters when computing the stochastic gradients and report improved generalization performance due to smoothing effects. However, this approach is poorly understood; it requires often model-specific noise and fine-tuning. To alleviate these drawbacks, we propose to use instead computationally efficient extrapolation (extragradient) to stabilize the optimization trajectory while still benefiting from smoothing to avoid sharp minima. This principled approach is well grounded from an optimization perspective and we show that a host of variations can be covered in a unified framework that we propose. We prove the convergence of this novel scheme and rigorously evaluate its empirical performance on ResNet, LSTM, and Transformer. We demonstrate that in a variety of experiments the scheme allows scaling to much larger batch sizes than before whilst reaching or surpassing SOTA accuracy.

The performance of optimizers, particularly in deep learning, depends considerably on their chosen hyperparameter configuration. The efficacy of optimizers is often studied under near-optimal problem-specific hyperparameters, and finding these settings may be prohibitively costly for practitioners. In this work, we argue that a fair assessment of optimizers’ performance must take the computational cost of hyperparameter tuning into account, i.e., how easy it is to find good hyperparameter configurations using an automatic hyperparameter search. Evaluating a variety of optimizers on an extensive set of standard datasets and architectures, our results indicate that Adam is the most practical solution, particularly in low-budget scenarios.

Federated Learning (FL) is a machine learning setting where many devices collaboratively train a machine learning model while keeping the training data decentralized. In most of the current training schemes the central model is refined by averaging the parameters of the server model and the updated parameters from the client side. However, directly averaging model parameters is only possible if all models have the same structure and size, which could be a restrictive constraint in many scenarios. In this work we investigate more powerful and more flexible aggregation schemes for FL. Specifically, we propose ensemble distillation for model fusion, i.e. training the central classifier through unlabeled data on the outputs of the models from the clients. This knowledge distillation technique mitigates privacy risk and cost to the same extent as the baseline FL algorithms, but allows flexible aggregation over heterogeneous client models that can differ e.g. in size, numerical precision or structure. We show in extensive empirical experiments on various CV/NLP datasets (CIFAR-10/100, ImageNet, AG News, SST2) and settings (heterogeneous models/data) that the server model can be trained much faster, requiring fewer communication rounds than any existing FL technique so far.

Combining different models is a widely used paradigm in machine learning applications. While the most common approach is to form an ensemble of models and average their individual predictions, this approach is often rendered infeasible by given resource constraints in terms of memory and computation, which grow linearly with the number of models. We present a layer-wise model fusion algorithm for neural networks that utilizes optimal transport to (soft-) align neurons across the models before averaging their associated parameters. We show that this can successfully yield "one-shot" knowledge transfer (i.e, without requiring any retraining) between neural networks trained on heterogeneous non-i.i.d. data. In both i.i.d. and non-i.i.d. settings, we illustrate that our approach significantly outperforms vanilla averaging, as well as how it can serve as an efficient replacement for the ensemble with moderate fine-tuning, for standard convolutional networks (like VGG11), residual networks (like ResNet18), and multi-layer perceptrons on CIFAR10, CIFAR100, and MNIST. Finally, our approach also provides a principled way to combine the parameters of neural networks with different widths, and we explore its application for model compression. The code is available at the following link, https://github.com/sidak/otfusion.

Lossy gradient compression has become a practical tool to overcome the communication bottleneck in centrally coordinated distributed training of machine learning models. However, algorithms for decentralized training with compressed communication over arbitrary connected networks have been more complicated, requiring additional memory and hyperparameters. We introduce a simple algorithm that directly compresses the model differences between neighboring workers using low-rank linear compressors. We prove that our method does not require any additional hyperparameters, converges faster than prior methods, and is asymptotically independent of both the network and the compression. Inspired the PowerSGD algorithm for centralized deep learning, we execute power iteration steps on model differences to maximize the information transferred per bit. Out of the box, these compressors perform on par with state-of-the-art tuned compression algorithms in a series of deep learning benchmarks.

Coordinate descent with random coordinate selection is the current state of the art for many large scale optimization problems. However, greedy selection of the steepest coordinate on smooth problems can yield convergence rates independent of the dimension $n$, requiring $n$ times fewer iterations. In this paper, we consider greedy updates that are based on subgradients for a class of non-smooth composite problems, including $L1$-regularized problems, SVMs and related applications. For these problems we provide (i) the first linear rates of convergence independent of $n$, and show that our greedy update rule provides speedups similar to those obtained in the smooth case. This was previously conjectured to be true for a stronger greedy coordinate selection strategy. Furthermore, we show that (ii) our new selection rule can be mapped to instances of maximum inner product search, allowing to leverage standard nearest neighbor algorithms to speed up the implementation. We demonstrate the validity of the approach through extensive numerical experiments.

A new generation of manycore processors is on the rise that offers dozens and more cores on a chip and, in a sense, fuses host processor and accelerator. In this paper we target the efficient training of generalized linear models on these machines. We propose a novel approach for achieving parallelism which we call Heterogeneous Tasks on Homogeneous Cores (HTHC). It divides the problem into multiple fundamentally different tasks, which themselves are parallelized. For evaluation, we design a detailed, architecture-cognizant implementation of our scheme on a recent 72-core Knights Landing processor that is adaptive to the cache, memory, and core structure. Our library efficiently supports dense and sparse datasets as well as 4-bit quantized data for further possible gains in performance. We show benchmarks for Lasso and SVM with different data sets against straightforward parallel implementations and prior software. In particular, for Lasso on dense data, we improve the state-of-the-art by an order of magnitude.

We identify a phenomenon, which we refer to as multi-model forgetting, that occurs when sequentially training multiple deep networks with partially-shared parameters; the performance of previously-trained models degrades as one optimizes a subsequent one, due to the overwriting of shared parameters. To overcome this, we introduce a statistically-justified weight plasticity loss that regularizes the learning of a model’s shared parameters according to their importance for the previous models, and demonstrate its effectiveness when training two models sequentially and for neural architecture search. Adding weight plasticity in neural architecture search preserves the best models to the end of the search and yields improved results in both natural language processing and computer vision tasks.

Sign-based algorithms (e.g. signSGD) have been proposed as a biased gradient compression technique to alleviate the communication bottleneck in training large neural networks across multiple workers. We show simple convex counter-examples where signSGD does not converge to the optimum. Further, even when it does converge, signSGD may generalize poorly when compared with SGD. These issues arise because of the biased nature of the sign compression operator. We then show that using error-feedback, i.e. incorporating the error made by the compression operator into the next step, overcomes these issues. We prove that our algorithm (EF-SGD) with arbitrary compression operator achieves the same rate of convergence as SGD without any additional assumptions. Thus EF-SGD achieves gradient compression for free. Our experiments thoroughly substantiate the theory.

We consider decentralized stochastic optimization with the objective function (e.g. data samples for machine learning tasks) being distributed over n machines that can only communicate to their neighbors on a fixed communication graph. To address the communication bottleneck, the nodes compress (e.g. quantize or sparsify) their model updates. We cover both unbiased and biased compression operators with quality denoted by \delta <= 1 (\delta=1 meaning no compression). We (i) propose a novel gossip-based stochastic gradient descent algorithm, CHOCO-SGD, that converges at rate O(1/(nT) + 1/(T \rho^2 \delta)^2) for strongly convex objectives, where T denotes the number of iterations and \rho the eigengap of the connectivity matrix. We (ii) present a novel gossip algorithm, CHOCO-GOSSIP, for the average consensus problem that converges in time O(1/(\rho^2\delta) \log (1/\epsilon)) for accuracy \epsilon > 0. This is (up to our knowledge) the first gossip algorithm that supports arbitrary compressed messages for \delta > 0 and still exhibits linear convergence. We (iii) show in experiments that both of our algorithms do outperform the respective state-of-the-art baselines and CHOCO-SGD can reduce communication by at least two orders of magnitudes.

Keyword Spotting (KWS) systems allow detecting a set of spoken (pre-defined) keywords. Open-vocabulary KWS systems search for the keywords in the set of word hypotheses generated by an automatic speech recognition (ASR) system which is computationally expensive and, therefore, often implemented as a cloud-based service. Besides, KWS systems could use also word classification algorithms that do not allow easily changing the set of words to be recognized, as the classes have to be defined a priori, even before training the system. In this paper, we propose the implementation of an open-vocabulary ASR-free KWS system based on speech and text encoders that allow matching the computed embeddings in order to spot whether a keyword has been uttered. This approach would allow choosing the set of keywords a posteriori while requiring low computational power. The experiments, performed on two different datasets, show that our method is competitive with other state of the art KWS systems while allowing for a flexibility of configuration and being computationally efficient.

Time series constitute a challenging data type for machine learning algorithms, due to their highly variable lengths and sparse labeling in practice. In this paper, we tackle this challenge by proposing an unsupervised method to learn universal embeddings of time series. Unlike previous works, it is scalable with respect to their length and we demonstrate the quality, transferability and practicability of the learned representations with thorough experiments and comparisons. To this end, we combine an encoder based on causal dilated convolutions with a novel triplet loss employing time-based negative sampling, obtaining general-purpose representations for variable length and multivariate time series.

We study gradient compression methods to alleviate the communication bottleneck in data-parallel distributed optimization. Despite the significant attention received, current compression schemes either do not scale well, or fail to achieve the target test accuracy. We propose a low-rank gradient compressor that can i) compress gradients rapidly, ii) efficiently aggregate the compressed gradients using all-reduce, and iii) achieve test performance on par with SGD. The proposed algorithm is the only method evaluated that achieves consistent wall-clock speedups when benchmarked against regular SGD with an optimized communication backend. We demonstrate reduced training times for convolutional networks as well as LSTMs on common datasets.

There has recently been much interest in extending vector-based word representations to multiple languages, such that words can be compared across languages. In this paper, we shift the focus from words to documents and introduce a method for embedding documents written in any language into a single, language-independent vector space. For training, our approach leverages a multilingual corpus where the same concept is covered in multiple languages (but not necessarily via exact translations), such as Wikipedia. Our method, Cr5 (Crosslingual reduced-rank ridge regression), starts by training a ridge-regression-based classifier that uses language-specific bag-of-word features in order to predict the concept that a given document is about. We show that, when constraining the learned weight matrix to be of low rank, it can be factored to obtain the desired mappings from language-specific bags-of-words to language-independent embeddings. As opposed to most prior methods, which use pretrained monolingual word vectors, postprocess them to make them crosslingual, and finally average word vectors to obtain document vectors, Cr5 is trained end-to-end and is thus natively crosslingual as well as document-level. Moreover, since our algorithm uses the singular value decomposition as its core operation, it is highly scalable. Experiments show that our method achieves state-of-the-art performance on a crosslingual document retrieval task. Finally, although not trained for embedding sentences and words, it also achieves competitive performance on crosslingual sentence and word retrieval tasks.

First-order optimization methods comprise two important primitives: i) the computation of gradient information and ii) the computation of the update that leads to the next iterate. In practice there is often a wide mismatch between the time required for the two steps, leading to underutilization of resources. In this work, we propose a new framework, Approx Composite Minimization (ACM) that uses approximate update steps to ensure balance between the two operations. The accuracy is adaptively chosen in an online fashion to take advantage of changing conditions. Our unified analysis for approximate composite minimization generalizes and extends previous work to new settings. Numerical experiments on Lasso regression and SVMs demonstrate the effectiveness of the novel scheme.

Keyphrase extraction is the task of automatically selecting a small set of phrases that best describe a given free text document.Supervised keyphrase extraction requires large amounts of labeled training data and generalizes very poorly outside the domain of the training data.At the same time, unsupervised systems have poor accuracy, and often do not generalize well, as they require the input document to belong to a larger corpus also given as input.Addressing these drawbacks, in this paper, we tackle keyphrase extraction from single documents with EmbedRank: a novel unsupervised method, that leverages sentence embeddings.EmbedRank achieves higher F-scores than graph-based state of the art systems on standard datasets and is suitable for real-time processing of large amounts of Web data.With EmbedRank, we also explicitly increase coverage and diversity among the selected keyphrases by introducing an embedding-based maximal marginal relevance (MMR) for new phrases.A user study including over 200 votes showed that, although reducing the phrases' semantic overlap leads to no gains in F-score, our high diversity selection is preferred by humans.

Due to the rapid growth of data and computational resources, distributed optimization has become an active research area in recent years. While first-order methods seem to dominate the field, second-order methods are nevertheless attractive as they potentially require fewer communication rounds to converge. However, there are significant drawbacks that impede their wide adoption, such as the computation and the communication of a large Hessian matrix. In this paper we present a new algorithm for distributed training of generalized linear models that only requires the computation of diagonal blocks of the Hessian matrix on the individual workers. To deal with this approximate information we propose an adaptive approach that - akin to trust-region methods - dynamically adapts the auxiliary model to compensate for modeling errors. We provide theoretical rates of convergence for a wide class of problems including $L_1$-regularized objectives. We also demonstrate that our approach achieves state-of-the-art results on multiple large benchmark datasets.

Two popular examples of first-order optimization methods over linear spaces are coordinate descent and matching pursuit algorithms, with their randomized variants. While the former targets the optimization by moving along coordinates, the latter considers a generalized notion of directions. Exploiting the connection between the two algorithms, we present a unified analysis of both, providing affine invariant sublinear $O(1/t)$ rates on smooth objectives and linear convergence on strongly convex objectives. As a byproduct of our affine invariant analysis of matching pursuit, our rates for steepest coordinate descent are the tightest known. Furthermore, we show the first accelerated convergence rate $O(1/t^2)$ for matching pursuit and steepest coordinate descent on convex objectives.

Matteo Pagliardini, Prakhar Gupta, Martin Jaggi. Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long Papers). 2018.

The wide adoption of DNNs has given birth to unrelenting computing requirements, forcing datacenter operators to adopt domain-specific accelerators to train them. These accelerators typically employ densely packed full-precision floating-point arithmetic to maximize performance per area. Ongoing research efforts seek to further increase that performance density by replacing floating-point with fixed-point arithmetic. However, a significant roadblock for these attempts has been fixed point's narrow dynamic range, which is insufficient for DNN training convergence. We identify block floating point (BFP) as a promising alternative representation since it exhibits wide dynamic range and enables the majority of DNN operations to be performed with fixed-point logic. Unfortunately, BFP alone introduces several limitations that preclude its direct applicability. In this work, we introduce HBFP, a hybrid BFP-FP approach, which performs all dot products in BFP and other operations in floating point. HBFP delivers the best of both worlds: the high accuracy of floating point at the superior hardware density of fixed point. For a wide variety of models, we show that HBFP matches floating point's accuracy while enabling hardware implementations that deliver up to 8.5x higher throughput.

Decentralized machine learning is a promising emerging paradigm in view of global challenges of data ownership and privacy. We consider learning of linear classification and regression models, in the setting where the training data is decentralized over many user devices, and the learning algorithm must run on-device, on an arbitrary communication network, without a central coordinator. We propose COLA, a new decentralized training algorithm with strong theoretical guarantees and superior practical performance. Our framework overcomes many limitations of existing methods, and achieves communication efficiency, scalability, elasticity as well as resilience to changes in data and allows for unreliable and heterogeneous participating devices.

Huge scale machine learning problems are nowadays tackled by distributed optimization algorithms, i.e. algorithms that leverage the compute power of many devices for training. The communication overhead is a key bottleneck that hinders perfect scalability. Various recent works proposed to use quantization or sparsification techniques to reduce the amount of data that needs to be communicated, for instance by only sending the most significant entries of the stochastic gradient (top-k sparsification). Whilst such schemes showed very promising performance in practice, they have eluded theoretical analysis so far. In this work we analyze Stochastic Gradient Descent (SGD) with k-sparsification or compression (for instance top-k or random-k) and show that this scheme converges at the same rate as vanilla SGD when equipped with error compensation (keeping track of accumulated errors in memory). That is, communication can be reduced by a factor of the dimension of the problem (sometimes even more) whilst still converging at the same rate. We present numerical experiments to illustrate the theoretical findings and the good scalability for distributed applications.

Two of the most fundamental prototypes of greedy optimization are the matching pursuit and Frank-Wolfe algorithms. In this paper we take a unified view on both classes of methods, leading to the first explicit convergence rates of matching pursuit methods in an optimization sense, for general sets of atoms. We derive sublinear (1/t) convergence for both classes on general smooth objectives, and linear convergence on strongly convex objectives, as well as a clear correspondence of algorithm variants. Our presented algorithms and rates are affine invariant, and do not need any incoherence or sparsity assumptions.

Coordinate descent methods employ random partial updates of decision variables in order to solve huge-scale convex optimization problems. In this work, we introduce new adaptive rules for the random selection of their updates. By adaptive, we mean that our selection rules are based on the dual residual or the primal-dual gap estimates and can change at each iteration. We theoretically characterize the performance of our selection rules and demonstrate improvements over the state-of-the-art, and extend our theory and algorithms to general convex objectives. Numerical evidence with hinge-loss support vector machines and Lasso confirm that the practice follows the theory.

We propose a new selection rule for the coordinate selection in coordinate descent methods for huge-scale optimization. The efficiency of this novel scheme is provably better than the efficiency of uniformly random selection, and can reach the efficiency of steepest coordinate descent (SCD), enabling an acceleration of a factor of up to $n$, the number of coordinates. In many practical applications, our scheme can be implemented at no extra cost and computational efficiency very close to the faster uniform selection. Numerical experiments with Lasso and Ridge regression show promising improvements, in line with our theoretical guarantees.

We propose a generic algorithmic building block to accelerate training of machine learning models on heterogeneous compute systems. Our scheme allows to efficiently employ compute accelerators such as GPUs and FPGAs for the training of large-scale machine learning models, when the training data exceeds their memory capacity. Also, it provides adaptivity to any system's memory hierarchy in terms of size and processing speed. Our technique is built upon novel theoretical insights regarding primal-dual coordinate methods, and uses duality gap information to dynamically decide which part of the data should be made available for fast processing. To illustrate the power of our approach we demonstrate its performance for training of generalized linear models on a large-scale dataset exceeding the memory size of a modern GPU, showing an order-of-magnitude speedup over existing approaches.

Greedy optimization methods such as Matching Pursuit (MP) and Frank-Wolfe (FW) algorithms regained popularity in recent years due to their simplicity, effectiveness and theoretical guarantees. MP and FW address optimization over the linear span and the convex hull of a set of atoms, respectively. In this paper, we consider the intermediate case of optimization over the convex cone, parametrized as the conic hull of a generic atom set, leading to the first principled definitions of non-negative MP algorithms for which we give explicit convergence rates and demonstrate excellent empirical performance. In particular, we derive sublinear (O(1/t)) convergence on general smooth and convex objectives, and linear convergence (O(e^{-t})) on strongly convex objectives, in both cases for general sets of atoms. Furthermore, we establish a clear correspondence of our algorithms to known algorithms from the MP and FW literature. Our novel algorithms and analyses target general atom sets and general objective functions, and hence are directly applicable to a large variety of learning settings.

Importance sampling has become an indispensable strategy to speed up optimization algorithms for large-scale applications. Improved adaptive variants -- using importance values defined by the complete gradient information which changes during optimization -- enjoy favorable theoretical properties, but are typically computationally infeasible. In this paper we propose an efficient approximation of gradient-based sampling, which is based on safe bounds on the gradient. The proposed sampling distribution is (i) provably the \emph{best sampling} with respect to the given bounds, (ii) always better than uniform sampling and fixed importance sampling and (iii) can efficiently be computed -- in many applications at negligible extra cost. The proposed sampling scheme is generic and can easily be integrated into existing algorithms. In particular, we show that coordinate-descent (CD) and stochastic gradient descent (SGD) can enjoy significant a speed-up under the novel scheme. The proven efficiency of the proposed sampling is verified by extensive numerical testing.

This paper presents a novel approach for multi-lingual sentiment classification in short texts. This is a challenging task as the amount of training data in languages other than English is very limited. Previously proposed multi-lingual approaches typically require to establish a correspondence to English for which powerful classifiers are already available. In contrast, our method does not require such supervision. We leverage large amounts of weakly-supervised data in various languages to train a multi-layer convolutional network and demonstrate the importance of using pre-training of such networks. We thoroughly evaluate our approach on various multi-lingual datasets, including the recent SemEval-2016 sentiment prediction benchmark (Task 4), where we achieved state-of-the-art performance. We also compare the performance of our model trained individually for each language to a variant trained for all languages at once. We show that the latter model reaches slightly worse - but still acceptable - performance when compared to the single language model, while benefiting from better generalization properties across languages.

This paper deals with semantic segmentation of high-resolution (aerial) images where a semantic class label is assigned to each pixel via supervised classification as a basis for automatic map generation. Recently, deep convolutional neural networks (CNNs) have shown impressive performance and have quickly become the de-facto standard for semantic segmentation, with the added benefit that task-specific feature design is no longer necessary. However, a major downside of deep learning methods is that they are extremely data hungry, thus aggravating the perennial bottleneck of supervised classification, to obtain enough annotated training data. On the other hand, it has been observed that they are rather robust against noise in the training labels. This opens up the intriguing possibility to avoid annotating huge amounts of training data, and instead train the classifier from existing legacy data or crowd-sourced maps that can exhibit high levels of noise. The question addressed in this paper is: can training with large-scale publicly available labels replace a substantial part of the manual labeling effort and still achieve sufficient performance? Such data will inevitably contain a significant portion of errors, but in return virtually unlimited quantities of it are available in larger parts of the world. We adapt a state-of-the-art CNN architecture for semantic segmentation of buildings and roads in aerial images, and compare its performance when using different training data sets, ranging from manually labeled pixel-accurate ground truth of the same city to automatic training data derived from OpenStreetMap data from distant locations. We report our results that indicate that satisfying performance can be obtained with significantly less manual annotation effort, by exploiting noisy large-scale training data.

We propose an algorithm-independent framework to equip existing optimization methods with primal-dual certificates. Such certificates and corresponding rate of convergence guarantees are important for practitioners to diagnose progress, in particular in machine learning applications. We obtain new primal-dual convergence rates, e.g., for the Lasso as well as many L1, Elastic Net, group Lasso and TV-regularized problems. The theory applies to any norm-regularized generalized linear model. Our approach provides efficiently computable duality gaps which are globally defined, without modifying the original problems in the region of interest.

Distributed optimization methods for large-scale machine learning suffer from a communication bottleneck. It is difficult to reduce this bottleneck while still efficiently and accurately aggregating partial work from different machines. In this paper, we present a novel generalization of the recent communication-efficient primal-dual framework (COCOA) for distributed optimization. Our framework, COCOA+, allows for additive combination of local updates to the global parameters at each iteration, whereas previous schemes only allow conservative averaging. We give stronger (primal-dual) convergence rate guarantees for both COCOA as well as our new variants, and generalize the theory for both methods to cover non-smooth convex loss functions. We provide an extensive experimental comparison that shows the markedly improved performance of COCOA+ on several real-world distributed datasets, especially when scaling up the number of machines.

The Frank-Wolfe (FW) optimization algorithm has lately re-gained popularity thanks in particular to its ability to nicely handle the structured constraints appearing in machine learning applications. However, its convergence rate is known to be slow (sublinear) when the solution lies at the boundary. A simple less-known fix is to add the possibility to take away steps' during optimization, an operation that importantly does not require a feasibility oracle. In this paper, we highlight and clarify several variants of the Frank-Wolfe optimization algorithm that has been successfully applied in practice: FW with away steps, pairwise FW, fully-corrective FW and Wolfe's minimum norm point algorithm, and prove for the first time that they all enjoy global linear convergence under a weaker condition than strong convexity. The constant in the convergence rate has an elegant interpretation as the product of the (classical) condition number of the function with a novel geometric quantity that plays the role of thecondition number' of the constraint set. We provide pointers to where these algorithms have made a difference in practice, in particular with the flow polytope, the marginal polytope and the base polytope for submodular optimization.

We provide stronger and more general primal-dual convergence results for Frank-Wolfe-type algorithms (a.k.a. conditional gradient) for constrained convex optimization, enabled by a simple framework of duality gap certificates. Our analysis also holds if the linear subproblems are only solved approximately (as well as if the gradients are inexact), and is proven to be worst-case optimal in the sparsity of the obtained solutions. On the application side, this allows us to unify a large variety of existing sparse greedy methods, in particular for optimization over convex hulls of an atomic set, even if those sets can only be approximated, including sparse (or structured sparse) vectors or matrices, low-rank matrices, permutation matrices, or max-norm bounded matrices. We present a new general framework for convex optimization over matrix factorizations, where every Frank-Wolfe iteration will consist of a low-rank update, and discuss the broad application areas of this approach.

We propose a randomized block-coordinate variant of the classic Frank-Wolfe algorithm for convex optimization with block-separable constraints. Despite its lower iteration cost, we show that it achieves a similar convergence rate in duality gap as the full Frank-Wolfe algorithm. We also show that, when applied to the dual structural support vector machine (SVM) objective, this yields an online algorithm that has the same low iteration complexity as primal stochastic subgradient methods. However, unlike stochastic subgradient methods, the block-coordinate Frank-Wolfe algorithm allows us to compute the optimal step-size and yields a computable duality gap guarantee. Our experiments indicate that this simple algorithm outperforms competing structural SVM solvers.

We consider parameterized convex optimization problems over the unit simplex, that depend on one parameter. We provide a simple and efficient scheme for maintaining an ϵ-approximate solution (and a corresponding ϵ-coreset) along the entire parameter path. We prove correctness and optimality of the method. Practically relevant instances of the abstract parameterized optimization problem are for example regularization paths of support vector machines, multiple kernel learning, and minimum enclosing balls of moving points.

Following recent work of Clarkson, we translate the coreset framework to the problems of finding the point closest to the origin inside a polytope, finding the shortest distance between two polytopes, Perceptrons, and soft- as well as hard-margin Support Vector Machines (SVM). We prove asymptotically matching upper and lower bounds on the size of coresets, stating that µ-coresets of size (1+o(1)) E*/µ do always exist as µ-0, and that this is best possible. The crucial quantity E* is what we call the excentricity of a polytope, or a pair of polytopes. Additionally, we prove linear convergence speed of Gilbert's algorithm, one of the earliest known approximation algorithms for polytope distance, and generalize both the algorithm and the proof to the two polytope case. Interestingly, our coreset bounds also imply that we can for the first time prove matching upper and lower bounds for the sparsity of Perceptron and SVM solutions.