Training networks to be compressible: low rank, quantization, and how they combine-Saclay-

Mathématiques, information  scientifique, logiciel Stage - Training networks to be compressible: low rank, quantization, and how they combine-Saclay-H/F To run on smaller hardware or serve more cheaply, large neural networks are often compressed before deployment: most commonly by quantization (storing numbers with few bits), less often by low-rank factorization (replacing a weight matrix by a product of two thin factors). Models are increasingly trained or fine-tuned so that they compress well, for instance by pushing their weights toward low rank. These methods are mostly judged on factorization alone, yet a factorized model is also quantized before it is deployed. Does preparing a model for low rank help or hurt once it is also quantized, and can the combination compete with quantization alone? Our first measurements suggest the answer is less simple than it looks. Our group develops compression methods that measure errors by their effect on the network's outputs, and that predict in closed form what rounding low-rank factors costs. As an intern at the CEA, you will have the opportunity to work in a world-renowned research environment. Our teams consist of passionate and dedicated experts, providing an environment conducive to learning and collaboration. You will have access to state-of-the-art equipment and top-tier research resources to carry out your assignments. The work performed may potentially lead to a scientific publication.Context:This internship aims to understand how training a network toward low rank changes what it costs to quantize its factors, using the team's closed-form price of rounding both to explain the effect and to act on it. It combines matrix analysis with controlled experiments on language models, to: What do we expect from you?Measure what low-rank preparation does to a model's compressibility, by factorization and by quantization, against an unprepared model and against quantization alone, at equal memory.Explain it: what the preparation changes in the factors, read through the closed-form price, and whether the price predicts the measured cost.Use the price to propose and test an improvement.The internship may lead to a PhD starting in October 2027.  #Cea List Profile:Master's (M2) or engineering-school student in machine learning, applied mathematics or computer scienceStrong linear algebra; probability and optimization are a plusSolid Python and PyTorchCare in designing experiments and reading their resultsExperience with Hugging Face language models is a plus Conformément aux engagements pris par le CEA en faveur de l'intégration des personnes handicapées, cet emploi est ouvert à toutes et à tous. Le CEA propose des aménagements et/ou des possibilités d'organisation pour l'inclusion des travailleurs handicapés.