Disentangling polysemantic neurons is at the core of many current approaches to interpretability of large language models. Here we attempt to study how disentanglement can be used to understand performance, particularly under weight sparsity, a leading post-training optimization technique. We suggest a novel measure for estimating neuronal entanglement: the Wasserstein distance of a neuron's output distribution to a Gaussian. Moreover, we show the existence of a small number of highly entangled "Wasserstein Neurons" in each linear layer of an LLM, characterized by their highly non-Gaussian output distributions, their role in mapping similar inputs to dissimilar outputs, and their significant impact on model accuracy. To study these phenomena, we propose a new experimental framework for disentangling polysemantic neurons. Our framework separates each layer's inputs to create a mixture of experts where each neuron's output is computed by a mixture of neurons of lower Wasserstein distance, each better at maintaining accuracy when sparsified without retraining. We provide strong evidence that this is because the mixture of sparse experts is effectively disentangling the input-output relationship of individual neurons, in particular the difficult Wasserstein neurons.

Modern deep neural networks exhibit heterogeneity across numerous layers of various types such as residuals, multi-head attention, etc., due to varying structures (dimensions, activation functions, etc.), distinct representation characteristics, which impact predictions. We develop a general layer-wise quantization framework with tight variance and code-length bounds, adapting to the heterogeneities over the course of training. We then apply a new layer-wise quantization technique within distributed variational inequalities (VIs), proposing a novel Quantized Optimistic Dual Averaging (QODA) algorithm with adaptive learning rates, which achieves competitive convergence rates for monotone VIs. We empirically show that QODA achieves up to a $150$% speedup over the baselines in end-to-end training time for training Wasserstein GAN on $12+$ GPUs.

Saleh Ashkboos, Ilia Markov, Elias Frantar, Tingxuan Zhong, Xincheng Wang, Jie Ren, Torsten Hoefler, Dan Alistarh. Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing. 2024.

Communication-reduction techniques are a popular way to improve scalability in data-parallel training of deep neural networks (DNNs). The recent emergence of large language models such as GPT has created the need for new approaches to exploit data-parallelism. Among these, fully-sharded data parallel (FSDP) training is highly popular, yet it still encounters scalability bottlenecks. One reason is that applying compression techniques to FSDP is challenging: as the vast majority of the communication involves the model’s weights, direct compression alters convergence and leads to accuracy loss. We present QSDP, a variant of FSDP which supports both gradient and weight quantization with theoretical guarantees, is simple to implement and has essentially no overheads. To derive QSDP we prove that a natural modification of SGD achieves convergence even when we only maintain quantized weights, and thus the domain over which we train consists of quantized points and is, therefore, highly non-convex. We validate this approach by training GPT-family models with up to 1.3 billion parameters on a multi-node cluster. Experiments show that QSDP preserves model accuracy, while completely removing the communication bottlenecks of FSDP, providing end-to-end speedups of up to 2.2x.

The ability to scale out training workloads has been one of the key performance enablers of deep learning. The main scaling approach is data-parallel GPU-based training, which has been boosted by hardware and software support for highly efficient point-to-point communication, and in particular via hardware bandwidth over-provisioning. Overprovisioning comes at a cost: there is an order of magnitude price difference between "cloud-grade" servers with such support, relative to their popular "consumer-grade" counterparts, although single server-grade and consumer-grade GPUs can have similar computational envelopes.

Abstract Native language identification (NLI)—the task of automatically identifying the native language (L1) of persons based on their writings in the second language (L2)—is based on the hypothesis that characteristics of L1 will surface and interfere in the production of texts in L2 to the extent that L1 is identifiable. We present an in-depth investigation of features that model a variety of linguistic phenomena potentially involved in native language interference in the context of the NLI task: the languages’ structuring of information through punctuation usage, emotion expression in language, and similarities of form with the L1 vocabulary through the use of anglicized words, cognates, and other misspellings. The results of experiments with different combinations of features in a variety of settings allow us to quantify the native language interference value of these linguistic phenomena and show how robust they are in cross-corpus experiments and with respect to proficiency in L2. These experiments provide a deeper insight into the NLI task, showing how native language interference explains the gap between baseline, corpus-independent features, and the state of the art that relies on features/representations that cover (indiscriminately) a variety of linguistic phenomena.

One key element behind the recent progress of machine learning has been the ability to train machine learning models in large-scale distributed shared-memory and message-passing environments. Most of these models are trained employing variants of stochastic gradient descent (SGD) based optimization, but most methods involve some type of consistency relaxation relative to sequential SGD, to mitigate its large communication or synchronization costs at scale. In this paper, we introduce a general consistency condition covering communication-reduced and asynchronous distributed SGD implementations. Our framework, called elastic consistency, decouples the system-specific aspects of the implementation from the SGD convergence requirements, giving a general way to obtain convergence bounds for a wide variety of distributed SGD methods used in practice. Elastic consistency can be used to re-derive or improve several previous convergence bounds in message-passing and shared-memory settings, but also to analyze new models and distribution schemes. As a direct application, we propose and analyze a new synchronization-avoiding scheduling scheme for distributed SGD, and show that it can be used to efficiently train deep convolutional models for image classification.

While messaging systems with strong security guarantees are widely used in practice, designing a protocol that scales efficiently to large groups and enjoys similar security guarantees remains largely open. The two existing proposals to date are ART (Cohn-Gordon et al., CCS18) and TreeKEM (IETF, The Messaging Layer Security Protocol, draft). TreeKEM is the currently considered candidate by the IETF MLS working group, but dynamic group operations (i.e. adding and removing users) can cause efficiency issues. In this paper we formalize and analyze a variant of TreeKEM which we term Tainted TreeKEM (TTKEM for short). The basic idea underlying TTKEM was suggested by Millican (MLS mailing list, February 2018). This version is more efficient than TreeKEM for some natural distributions of group operations, we quantify this through simulations.Our second contribution is two security proofs for TTKEM which establish post compromise and forward secrecy even against adaptive attackers. The security loss (to the underlying PKE) in the Random Oracle Model is a polynomial factor, and a quasipolynomial one in the Standard Model. Our proofs can be adapted to TreeKEM as well. Before our work no security proof for any TreeKEM-like protocol establishing tight security against an adversary who can adaptively choose the sequence of operations was known. We also are the first to prove (or even formalize) active security where the server can arbitrarily deviate from the protocol specification. Proving fully active security – where also the users can arbitrarily deviate – remains open.

Native language identification (NLI) -identifying the native language (L1) of a person based on his/her writing in the second language (L2) -is useful for a variety of purposes, including marketing, security, and educational applications.From a traditional machine learning perspective, NLI is usually framed as a multi-class classification task, where numerous designed features are combined in order to achieve state-of-the-art results.We introduce a deep generative language modelling (LM) approach to NLI, which consists in fine-tuning a GPT-2 model separately on texts written by the authors with the same L1, and assigning a label to an unseen text based on the minimum LM loss with respect to one of these fine-tuned GPT-2 models.Our method outperforms traditional machine learning approaches and currently achieves the best results on the benchmark NLI datasets.

Many communication-efficient variants of SGD use gradient quantization schemes. These schemes are often heuristic and fixed over the course of training. We empirically observe that the statistics of gradients of deep models change during the training. Motivated by this observation, we introduce two adaptive quantization schemes, ALQ and AMQ. In both schemes, processors update their compression schemes in parallel by efficiently computing sufficient statistics of a parametric distribution. We improve the validation accuracy by almost 2% on CIFAR-10 and 1% on ImageNet in challenging low-cost communication setups. Our adaptive methods are also significantly more robust to the choice of hyperparameters.

In this paper, we describe experiments designed to explore and evaluate the impact of punctuation marks on the task of native language identification. Punctuation is specific to each language, and is part of the indicators that overtly represent the manner in which each language organizes and conveys information. Our experiments are organized in various set-ups: the usual multi-class classification for individual languages, also considering classification by language groups, across different proficiency levels, topics and even cross-corpus. The results support our hypothesis that punctuation marks are persistent and robust indicators of the native language of the author, which do not diminish in influence even when a high proficiency level in a non-native language is achieved.