Martin Jaggi

Zeming Chen, Alejandro Hernández Cano, Angelika Romanou et al. · allen-ai

Large language models (LLMs) can potentially democratize access to medical knowledge. While many efforts have been made to harness and improve LLMs' medical knowledge and reasoning capacities, the resulting models are either closed-source (e.g., PaLM, GPT-4) or limited in scale (<= 13B parameters), which restricts their abilities. In this work, we improve access to large-scale medical LLMs by releasing MEDITRON: a suite of open-source LLMs with 7B and 70B parameters adapted to the medical domain. MEDITRON builds on Llama-2 (through our adaptation of Nvidia's Megatron-LM distributed trainer), and extends pretraining on a comprehensively curated medical corpus, including selected PubMed articles, abstracts, and internationally-recognized medical guidelines. Evaluations using four major medical benchmarks show significant performance gains over several state-of-the-art baselines before and after task-specific finetuning. Overall, MEDITRON achieves a 6% absolute performance gain over the best public baseline in its parameter class and 3% over the strongest baseline we finetuned from Llama-2. Compared to closed-source LLMs, MEDITRON-70B outperforms GPT-3.5 and Med-PaLM and is within 5% of GPT-4 and 10% of Med-PaLM-2. We release our code for curating the medical pretraining corpus and the MEDITRON model weights to drive open-source development of more capable medical LLMs.

Jean Ogier du Terrail, Samy-Safwan Ayed, Edwige Cyffers et al. · eth-zurich

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}.

Thijs Vogels, Hadrien Hendrikx, Martin Jaggi

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. This paper is an extension of the conference paper by Vogels et. al. (2022). Code: https://github.com/epfml/topology-in-decentralized-learning.

Anastasia Koloskova, Sebastian U. Stich, Martin Jaggi

We study the asynchronous stochastic gradient descent algorithm for distributed training over $n$ workers which have varying computation and communication frequency over time. In this algorithm, workers compute stochastic gradients in parallel at their own pace and return those to the server without any synchronization. Existing convergence rates of this algorithm for non-convex smooth objectives depend on the maximum gradient delay $τ_{\max}$ and show that an $ε$-stationary point is reached after $\mathcal{O}\!\left(σ^2ε^{-2}+ τ_{\max}ε^{-1}\right)$ iterations, where $σ$ denotes the variance of stochastic gradients. In this work (i) we obtain a tighter convergence rate of $\mathcal{O}\!\left(σ^2ε^{-2}+ \sqrt{τ_{\max}τ_{avg}}ε^{-1}\right)$ without any change in the algorithm where $τ_{avg}$ is the average delay, which can be significantly smaller than $τ_{\max}$. We also provide (ii) a simple delay-adaptive learning rate scheme, under which asynchronous SGD achieves a convergence rate of $\mathcal{O}\!\left(σ^2ε^{-2}+ τ_{avg}ε^{-1}\right)$, 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. In particular, we show that the heterogeneity term in convergence rate is only affected by the average delay within each worker.

Fedor Moiseev, Zhe Dong, Enrique Alfonseca et al.

Large language models (LLMs) have demonstrated human-level performance on a vast spectrum of natural language tasks. However, it is largely unexplored whether they can better internalize knowledge from a structured data, such as a knowledge graph, or from text. In this work, we propose a method to infuse structured knowledge into LLMs, by directly training T5 models on factual triples of knowledge graphs (KGs). We show that models pre-trained on Wikidata KG with our method outperform the T5 baselines on FreebaseQA and WikiHop, as well as the Wikidata-answerable subset of TriviaQA and NaturalQuestions. The models pre-trained on factual triples compare competitively with the ones on natural language sentences that contain the same knowledge. Trained on a smaller size KG, WikiMovies, we saw 3x improvement of exact match score on MetaQA task compared to T5 baseline. The proposed method has an advantage that no alignment between the knowledge graph and text corpus is required in curating training data. This makes our method particularly useful when working with industry-scale knowledge graphs.

Simin Fan, Matteo Pagliardini, Martin Jaggi

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 of (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 learned 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 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, and consistently achieves better test perplexity on the target domain.

Mariel Werner, Lie He, Michael Jordan et al.

Identifying clients with similar objectives and learning a model-per-cluster is an intuitive and interpretable approach to personalization in federated learning. However, doing so with provable and optimal guarantees has remained an open challenge. We formalize this problem as a stochastic optimization problem, achieving optimal convergence rates for a large class of loss functions. We propose simple iterative algorithms which identify clusters of similar clients and train a personalized model-per-cluster, using local client gradients and flexible constraints on the clusters. The convergence rates of our algorithms asymptotically match those obtained if we knew the true underlying clustering of the clients and are provably robust in the Byzantine setting where some fraction of the clients are malicious.

Thijs Vogels, Hadrien Hendrikx, Martin Jaggi

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.

Martin Jaggi

For the general problem of minimizing a convex function over a compact convex domain, we will investigate a simple iterative approximation algorithm based on the method by Frank & Wolfe 1956, that does not need projection steps in order to stay inside the optimization domain. Instead of a projection step, the linearized problem defined by a current subgradient is solved, which gives a step direction that will naturally stay in the domain. Our framework generalizes the sparse greedy algorithm of Frank & Wolfe and its primal-dual analysis by Clarkson 2010 (and the low-rank SDP approach by Hazan 2008) to arbitrary convex domains. We give a convergence proof guaranteeing ε-small duality gap after O(1/ε) iterations. The method allows us to understand the sparsity of approximate solutions for any l1-regularized convex optimization problem (and for optimization over the simplex), expressed as a function of the approximation quality. We obtain matching upper and lower bounds of Θ(1/ε) for the sparsity for l1-problems. The same bounds apply to low-rank semidefinite optimization with bounded trace, showing that rank O(1/ε) is best possible here as well. As another application, we obtain sparse matrices of O(1/ε) non-zero entries as ε-approximate solutions when optimizing any convex function over a class of diagonally dominant symmetric matrices. We show that our proposed first-order method also applies to nuclear norm and max-norm matrix optimization problems. For nuclear norm regularized optimization, such as matrix completion and low-rank recovery, we demonstrate the practical efficiency and scalability of our algorithm for large matrix problems, as e.g. the Netflix dataset. For general convex optimization over bounded matrix max-norm, our algorithm is the first with a convergence guarantee, to the best of our knowledge.

Matteo Pagliardini, Daniele Paliotta, Martin Jaggi et al.

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 implementations concerns and end up imposing a simple and static structure over the attention matrix. Conversely, implementing more dynamic sparse attentions 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.

Nikita Doikov, El Mahdi Chayti, Martin Jaggi

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 arithmetical 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 arithmetical complexity of second-order algorithms by a factor $\sqrt{d}$.

Yatin Dandi, Anastasia Koloskova, Martin Jaggi et al.

Decentralized learning provides an effective framework to train machine learning models with data distributed over arbitrary communication graphs. However, most existing approaches toward decentralized learning disregard the interaction between data heterogeneity and graph topology. In this paper, we characterize the dependence of convergence on the relationship between the mixing weights of the graph and the data heterogeneity across nodes. We propose a metric that quantifies the ability of a graph to mix the current gradients. We further prove that the metric controls the convergence rate, particularly in settings where the heterogeneity across nodes dominates the stochasticity between updates for a given node. Motivated by our analysis, we propose an approach that periodically and efficiently optimizes the metric using standard convex constrained optimization and sketching techniques. Through comprehensive experiments on standard computer vision and NLP benchmarks, we show that our approach leads to improvement in test performance for a wide range of tasks.

Vinitra Swamy, Malika Satayeva, Jibril Frej et al.

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.

El Mahdi Chayti, Nikita Doikov, Martin Jaggi

We study stochastic Cubic Newton methods for solving general possibly non-convex minimization problems. We propose a new framework, which we call the helper framework, that provides a unified view of the stochastic and variance-reduced second-order algorithms equipped with global complexity guarantees. It can also be applied to learning with auxiliary information. Our helper framework offers the algorithm designer high flexibility for constructing and analyzing the stochastic Cubic Newton methods, allowing arbitrary size batches, and the use of noisy and possibly biased estimates of the gradients and Hessians, incorporating both the variance reduction and the lazy Hessian updates. We recover the best-known complexities for the stochastic and variance-reduced Cubic Newton, under weak assumptions on the noise. A direct consequence of our theory is the new lazy stochastic second-order method, which significantly improves the arithmetic complexity for large dimension problems. We also establish complexity bounds for the classes of gradient-dominated objectives, that include convex and strongly convex problems. For Auxiliary Learning, we show that using a helper (auxiliary function) can outperform training alone if a given similarity measure is small.

Amirkeivan Mohtashami, Martin Jaggi, Sebastian Stich

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.

Simla Burcu Harma, Ayan Chakraborty, Nicholas Sperry et al.

The unprecedented demand for computing resources to train DNN models has led to a search for minimal numerical encoding. Recent state-of-the-art (SOTA) proposals advocate for multi-level scaled narrow bitwidth numerical formats. In this paper, we show that single-level scaling is sufficient to maintain training accuracy while maximizing arithmetic density. We identify a previously proposed single-level scaled format for 8-bit training, Hybrid Block Floating Point (HBFP), as the optimal candidate to minimize. We perform a full-scale exploration of the HBFP design space using mathematical tools to study the interplay among various parameters and identify opportunities for even smaller encodings across layers and epochs. Based on our findings, we propose Accuracy Booster, a mixed-mantissa HBFP technique that uses 4-bit mantissas for over 99% of all arithmetic operations in training and 6-bit mantissas only in the last epoch and first/last layers. We show Accuracy Booster enables increasing arithmetic density over all other SOTA formats by at least 2.3x while achieving state-of-the-art accuracies in 4-bit training.

Maria-Luiza Vladarean, Nikita Doikov, Martin Jaggi et al.

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.

Frédéric Berdoz, Abhishek Singh, Martin Jaggi et al.

Privacy-preserving machine learning has become a key conundrum for multi-party artificial intelligence. Federated learning (FL) and Split Learning (SL) are two frameworks that enable collaborative learning while keeping the data private (on device). In FL, each data holder trains a model locally and releases it to a central server for aggregation. In SL, the clients must release individual cut-layer activations (smashed data) to the server and wait for its response (during both inference and back propagation). While relevant in several settings, both of these schemes have a high communication cost, rely on server-level computation algorithms and do not allow for tunable levels of collaboration. In this work, we present a novel approach for privacy-preserving machine learning, where the clients collaborate via online knowledge distillation using a contrastive loss (contrastive w.r.t. the labels). The goal is to ensure that the participants learn similar features on similar classes without sharing their input data. To do so, each client releases averaged last hidden layer activations of similar labels to a central server that only acts as a relay (i.e., is not involved in the training or aggregation of the models). Then, the clients download these last layer activations (feature representations) of the ensemble of users and distill their knowledge in their personal model using a contrastive objective. For cross-device applications (i.e., small local datasets and limited computational capacity), this approach increases the utility of the models compared to independent learning and other federated knowledge distillation (FD) schemes, is communication efficient and is scalable with the number of clients. We prove theoretically that our framework is well-posed, and we benchmark its performance against standard FD and FL on various datasets using different model architectures.

Saleh Ashkboos, Amirkeivan Mohtashami, Maximilian L. Croci et al.

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 LLaMa2 models without any calibration data using round-to-nearest quantization. Code is available at: https://github.com/spcl/QuaRot.

Amirkeivan Mohtashami, Matteo Pagliardini, Martin Jaggi

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.

Simin Fan, Martin Jaggi

Automatic data selection and curriculum design for training large language models is challenging, with only a few existing methods showing improvements over standard training. Furthermore, current schemes focus on domain-level selection, overlooking the more fine-grained contributions of each individual training point. It is difficult to apply traditional datapoint selection methods on large language models: most online batch selection methods perform two-times forward or backward passes, which introduces considerable extra costs with large-scale models. To mitigate these obstacles, we propose irreducible curriculum as a curriculum learning algorithm for language model pretraining, which prioritizes samples with higher learnability. Specifically, to avoid prohibitive extra computation overhead, we simulate the sample loss along the main model's training trajectory using a small-scale proxy model. Our experiments on the RedPajama-1B dataset demonstrate a consistent improvement on validation perplexity across all 7 domains compared to random uniform baseline and the anti-curriculum strategy. Our method also reduces the sharpness of the network and illustrates a better 5-shot accuracy on MMLU benchmarks.

Cécile Trottet, Thijs Vogels, Martin Jaggi et al.

Data-driven Clinical Decision Support Systems (CDSS) have the potential to improve and standardise care with personalised probabilistic guidance. However, the size of data required necessitates collaborative learning from analogous CDSS's, which are often unsharable or imperfectly interoperable (IIO), meaning their feature sets are not perfectly overlapping. We propose Modular Clinical Decision Support Networks (MoDN) which allow flexible, privacy-preserving learning across IIO datasets, while providing interpretable, continuous predictive feedback to the clinician. MoDN is a novel decision tree composed of feature-specific neural network modules. It creates dynamic personalised representations of patients, and can make multiple predictions of diagnoses, updatable at each step of a consultation. The modular design allows it to compartmentalise training updates to specific features and collaboratively learn between IIO datasets without sharing any data.

Dongyang Fan, Diba Hashemi, Sai Praneeth Karimireddy et al.

Incorporating metadata in Large Language Models (LLMs) pretraining has recently emerged as a promising approach to accelerate training. However prior work highlighted only one useful signal-URLs, leaving open the question of whether other forms of metadata could yield greater benefits. In this study, we investigate a wider range of metadata types and find other types of metadata, such as fine-grained indicators of document quality that can also accelerate pretraining when prepended. We identify a common feature among effective metadata: they encode information at a finer granularity. We further introduce metadata appending as a means of improving training efficiency, where predicting an appropriate metadata as auxiliary task can help speed up pretraining. In addition, learnable meta-tokens trained with masked loss can recover part of the speedup by inducing quality-aware latent structure. Using probing, we analyze latent representations to understand how metadata shapes learning. Together, these results yield practical guidelines for integrating metadata to improve both the efficiency and effectiveness of LLM pretraining.

Ashok Vardhan Makkuva, Marco Bondaschi, Thijs Vogels et al.

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: LineAr CompreSsion in WirEless DistRibuted 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.

El Mahdi Chayti, Martin Jaggi

Learning new tasks by drawing on prior experience gathered from other (related) tasks is a core property of any intelligent system. Gradient-based meta-learning, especially MAML and its variants, has emerged as a viable solution to accomplish this goal. One problem MAML encounters is its computational and memory burdens needed to compute the meta-gradients. We propose a new first-order variant of MAML that we prove converges to a stationary point of the MAML objective, unlike other first-order variants. We also show that the MAML objective does not satisfy the smoothness assumption assumed in previous works; we show instead that its smoothness constant grows with the norm of the meta-gradient, which theoretically suggests the use of normalized or clipped-gradient methods compared to the plain gradient method used in previous works. We validate our theory on a synthetic experiment.

Ashok Vardhan Makkuva, Marco Bondaschi, Adway Girish et al.

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. Code is available at https://github.com/Bond1995/Markov .

Seyed Ali Bahrainian, Martin Jaggi, Carsten Eickhoff

Controlled abstractive summarization focuses on producing condensed versions of a source article to cover specific aspects by shifting the distribution of generated text towards a desired style, e.g., a set of topics. Subsequently, the resulting summaries may be tailored to user-defined requirements. This paper presents a new Transformer-based architecture capable of producing topic-focused summaries. The architecture modifies the cross-attention mechanism of the Transformer to bring topic-focus control to the generation process while not adding any further parameters to the model. We show that our model sets a new state of the art on the NEWTS dataset in terms of topic-focused abstractive summarization as well as a topic-prevalence score. Moreover, we show via extensive experiments that our proposed topical cross-attention mechanism can be plugged into various Transformer models, such as BART and T5, improving their performance on the CNN/Dailymail and XSum benchmark datasets for abstractive summarization. This is achieved via fine-tuning, without requiring training from scratch. Finally, we show through human evaluation that our model generates more faithful summaries outperforming the state-of-the-art Frost model.

Andrei Semenov, Matteo Pagliardini, Martin Jaggi

The recent development of Large Language Models (LLMs) has been accompanied by an effervescence of novel ideas and methods to better optimize the loss of deep learning models. Claims from those methods are myriad: from faster convergence to removing reliance on certain hyperparameters. However, the diverse experimental protocols used to validate these claims make direct comparisons between methods challenging. This study presents a comprehensive evaluation of recent optimization techniques across standardized LLM pretraining scenarios, systematically varying model size, batch size, and training duration. Through careful tuning of each method, we provide guidance to practitioners on which optimizer is best suited for each scenario. For researchers, our work highlights promising directions for future optimization research. Finally, by releasing our code and making all experiments fully reproducible, we hope our efforts can help the development and rigorous benchmarking of future methods.

Robin Faro, Dongyang Fan, Tamar Alphaidze et al.

Large language models (LLMs) are typically trained on fixed snapshots of the web, which means that their knowledge becomes stale and their predictions risk temporal leakage: relying on information that lies in the future relative to a query. We tackle this problem by pre-training from scratch a set of GPT-style experts on disjoint two-year slices of a 2013-2024 corpus and combining them through TiMoE, a Time-aware Mixture of Language Experts. At inference time, TiMoE masks all experts whose training window ends after the query timestamp and merges the remaining log-probabilities in a shared space, guaranteeing strict causal validity while retaining the breadth of multi-period knowledge. We also release TSQA, a 10k-question benchmark whose alternatives are explicitly labelled as past, future or irrelevant, allowing fine-grained measurement of temporal hallucinations. Experiments on eight standard NLP tasks plus TSQA show that a co-adapted TiMoE variant matches or exceeds the best single-period expert and cuts future-knowledge errors by up to 15%. Our results demonstrate that modular, time-segmented pre-training paired with causal routing is a simple yet effective path toward LLMs that stay chronologically grounded without sacrificing general performance much. We open source our code at TiMoE (Github): https://github.com/epfml/TiMoE

Amirkeivan Mohtashami, Martin Jaggi

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/.

Thijs Vogels, Lie He, Anastasia Koloskova et al.

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. Our code is available at http://github.com/epfml/relaysgd.

Sai Praneeth Karimireddy, Lie He, Martin Jaggi

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. Our code is open sourced at https://github.com/epfml/byzantine-robust-optimizer.

Sidak Pal Singh, Martin Jaggi

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.

Thijs Vogels, Sai Praneeth Karimireddy, Martin Jaggi

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 new low-rank gradient compressor based on power iteration 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. Our code is available at https://github.com/epfml/powersgd.

Sai Praneeth Karimireddy, Quentin Rebjock, Sebastian U. Stich et al.

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 and show that error-feedback improves both convergence and generalization. Code can be found at \url{https://github.com/epfml/error-feedback-SGD}.

Sidak Pal Singh, Andreas Hug, Aymeric Dieuleveut et al.

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 prebuilt histograms, are available under https://github.com/context-mover/.

Guilherme Penedo, Hynek Kydlíček, Vinko Sabolčec et al. · huggingface

Pre-training state-of-the-art large language models (LLMs) requires vast amounts of clean and diverse text data. While the open development of large high-quality English pre-training datasets has seen substantial recent progress, training performant multilingual LLMs remains a challenge, in large part due to the inherent difficulty of tailoring filtering and deduplication pipelines to a large number of languages. In this work, we introduce a new pre-training dataset curation pipeline based on FineWeb that can be automatically adapted to support any language. We extensively ablate our pipeline design choices on a set of nine diverse languages, guided by a set of meaningful and informative evaluation tasks that were chosen through a novel selection process based on measurable criteria. Ultimately, we show that our pipeline can be used to create non-English corpora that produce more performant models than prior datasets. We additionally introduce a straightforward and principled approach to rebalance datasets that takes into consideration both duplication count and quality, providing an additional performance uplift. Finally, we scale our pipeline to over 1000 languages using almost 100 Common Crawl snapshots to produce FineWeb2, a new 20 terabyte (5 billion document) multilingual dataset which we release along with our pipeline, training, and evaluation codebases.

Atli Kosson, Bettina Messmer, Martin Jaggi

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 $Δ\mathbf{w}_t = η_t \mathbf{u}_t$ early in training by using lower values for the learning rate $η_t$. In this work we argue that warmup benefits training by keeping the overall size of $Δ\mathbf{w}_t$ 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.

Youssef Allouah, Anastasia Koloskova, Aymane El Firdoussi et al.

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.

Dongyang Fan, Bettina Messmer, Martin Jaggi

In this study, we systematically evaluate the impact of common design choices in Mixture of Experts (MoEs) on validation performance, uncovering distinct influences at token and sequence levels. We also present empirical evidence showing comparable performance between a learned router and a frozen, randomly initialized router, suggesting that learned routing may not be essential. Our study further reveals that Sequence-level routing can result in topic-specific weak expert specialization, in contrast to syntax specialization observed with Token-level routing.

Matteo Pagliardini, Amirkeivan Mohtashami, Francois Fleuret et al.

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.

Vinitra Swamy, Syrielle Montariol, Julian Blackwell et al.

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.

Bettina Messmer, Vinko Sabolčec, Martin Jaggi

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 propose 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. 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.

Nicolas Wagner, Dongyang Fan, Martin Jaggi

We explore on-device self-supervised collaborative fine-tuning of large language models with limited local data availability. Taking inspiration from the collaborative learning community, we introduce three distinct trust-weighted gradient aggregation schemes: weight similarity-based, prediction similarity-based and validation performance-based. To minimize communication overhead, we integrate Low-Rank Adaptation (LoRA) and only exchange LoRA weight updates. Our protocols, driven by prediction and performance metrics, surpass both FedAvg and local fine-tuning methods, which is particularly evident in realistic scenarios with more diverse local data distributions. The results underscore the effectiveness of our approach in addressing heterogeneity and scarcity within local datasets.

Ramón Calvo González, Daniele Paliotta, Matteo Pagliardini et al.

Large Language Models (LLMs) demonstrate remarkable capabilities at the cost of high compute requirements. Recent studies have demonstrated that intermediate layers in LLMs can be removed or reordered without substantial accuracy loss; however, this insight has not yet been exploited to improve inference efficiency. Leveraging observed layer independence, we propose a novel method that groups consecutive layers into pairs evaluated in parallel, effectively restructuring the computational graph to enhance parallelism. Without requiring retraining or fine-tuning, this approach achieves an inference throughput improvement of 1.05x-1.20x on standard benchmarks, retaining 95\%-99\% of the original model accuracy. Empirical results demonstrate the practicality of this method in significantly reducing inference cost for large-scale LLM deployment. Additionally, we demonstrate that modest performance degradation can be substantially mitigated through lightweight fine-tuning, further enhancing the method's applicability.

El Mahdi Chayti, Nikita Doikov, Martin Jaggi

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.

Dongyang Fan, Vinko Sabolčec, Matin Ansaripour et al.

The increasing adoption of web crawling opt-outs by copyright holders of online content raises critical questions about the impact of data compliance on large language model (LLM) performance. However, little is known about how these restrictions (and the resultant filtering of pretraining datasets) affect the capabilities of models trained using these corpora. In this work, we conceptualize this effect as the $\textit{data compliance gap}$ (DCG), which quantifies the performance difference between models trained on datasets that comply with web crawling opt-outs, and those that do not. We measure the data compliance gap in two settings: pretraining models from scratch and continual pretraining from existing compliant models (simulating a setting where copyrighted data could be integrated later in pretraining). Our experiments with 1.5B models show that, as of January 2025, compliance with web data opt-outs does not degrade general knowledge acquisition (close to 0\% DCG). However, in specialized domains such as biomedical research, excluding major publishers leads to performance declines. These findings suggest that while general-purpose LLMs can be trained to perform equally well using fully open data, performance in specialized domains may benefit from access to high-quality copyrighted sources later in training. Our study provides empirical insights into the long-debated trade-off between data compliance and downstream model performance, informing future discussions on AI training practices and policy decisions. Our website is available at https://data-compliance.github.io/.

Andrei Semenov, Martin Jaggi, Nikita Doikov

In this work, we develop new optimization algorithms that use approximate second-order information combined with the gradient regularization technique to achieve fast global convergence rates for both convex and non-convex objectives. The key innovation of our analysis is a novel notion called Gradient-Normalized Smoothness, which characterizes the maximum radius of a ball around the current point that yields a good relative approximation of the gradient field. Our theory establishes a natural intrinsic connection between Hessian approximation and the linearization of the gradient. Importantly, Gradient-Normalized Smoothness does not depend on the specific problem class of the objective functions, while effectively translating local information about the gradient field and Hessian approximation into the global behavior of the method. This new concept equips approximate second-order algorithms with universal global convergence guarantees, recovering state-of-the-art rates for functions with Hölder-continuous Hessians and third derivatives, quasi-self-concordant functions, as well as smooth classes in first-order optimization. These rates are achieved automatically and extend to broader classes, such as generalized self-concordant functions. We demonstrate direct applications of our results for global linear rates in logistic regression and softmax problems with approximate Hessians, as well as in non-convex optimization using Fisher and Gauss-Newton approximations.

Atli Kosson, Jeremy Welborn, Yang Liu et al.

Transferring the optimal learning rate from small to large neural networks can enable efficient training at scales where hyperparameter tuning is otherwise prohibitively expensive. To this end, the Maximal Update Parameterization (muP) proposes a learning rate scaling designed to keep the update dynamics of internal representations stable across different model widths. However, the scaling rules of muP rely on strong assumptions, particularly about the geometric alignment of a layer's inputs with both its weights and gradient updates. In this large-scale empirical investigation, we show that these assumptions hold only briefly at the start of training in the practical setups where learning rate transfer is most valuable, such as LLM training. For the remainder of training it is weight decay rather than muP that correctly stabilizes the update dynamics of internal representations across widths, facilitating learning rate transfer. This suggests muP's scaling primarily acts as a form of implicit learning rate warmup, allowing us to largely replace it with modified warmup schedules. Together these findings fundamentally challenge prevailing beliefs about learning rate transfer and can explain empirical practice such as why muP requires the independent weight decay variant for successful transfer.

El Mahdi Chayti, Martin Jaggi

Stochastic difference-of-convex (DC) optimization is prevalent in numerous machine learning applications, yet its convergence properties under small batch sizes remain poorly understood. Existing methods typically require large batches or strong noise assumptions, which limit their practical use. In this work, we show that momentum enables convergence under standard smoothness and bounded variance assumptions (of the concave part) for any batch size. We prove that without momentum, convergence may fail regardless of stepsize, highlighting its necessity. Our momentum-based algorithm achieves provable convergence and demonstrates strong empirical performance.