Robindar

David A. R. Robin

David A. R. Robin's picture

Postdoc at Dauphine University

Machine Learning Phd

Publications

  1. Stab-SGD: Noise-Adaptivity in Smooth Optimization with Stability Ratios

    NeurIPS 25, Neural Information Processing Systems, San Diego, 2025

    A variant of SGD computing the stability ratio (relative noise level) of gradient estimates to automatically compute a scheduler to shrink step-sizes, with proofs of adaptivity in expectated loss last-iterate loss values, matching nearly all best rates of SGD with noise-tuned schedulers.

    Full text :  [ OpenReview ] 

  2. Random Sparse Lifts: Con­struc­tion, Ana­ly­sis and Con­ver­gence of fi­ni­te sparse net­works

    ICLR 24, International Conference on Learning Representations, Vienna, 2024

    Proof of convergence of finite-width multi-layer networks (and transformer-likes) to arbitrarily low loss values by gradient flow, when initialization is diverse and sparse enough. This shows that Probable-Approximate-Correctness is a type of structural guarantee that is achievable for large neural networks of essentially any architecture.

    Full text :  [ OpenReview ] 

  3. Con­ver­gence be­yond the over­pa­ram­et­er­ized re­gi­me with Ray­leigh quot­ients

    NeurIPS 22, Neural Information Processing Systems, New Orleans, 2022

    Proof of convergence of two-layer neural networks of finite width to arbitrarily low loss values under gradient flow. Without over-parameterization assumptions and thus stronger than infinite-width simplifications, this is achieved by integration of Kurdyka-Lojasiewicz inequalities, a technique to show optimal convergence even without convexity.

    Full text & code :  [ OpenReview ]  [ Github ] 

  4. Pe­rio­dic Sig­nal Re­covery with Re­gu­la­rized Sine Neu­ral Net­works

    NeurIPS 22, Neural Information Processing Systems, Neur­Reps Work­shop, New Orleans, 2022

    Neural networks fail to learn periodic functions of unknown frequency, even with sine-like activations, despite previously claimed fixes. Obstructions identified include need for a more diverse (non-vanishing high-variance) init and non-convex sparsity-promoting regularization. With both, perfect recovery far outside the training interval.

    Full text & code :  [ OpenReview ]  [ Github ] 

  5. Attacking and Fixing the Android Protected Confirmation Protocol

    Euro S&P 25, IEEE European Symposium on Security and Privacy, Venice, 2025

    Android's Protected Confirmation (APC) protocol exhibits two vulnerabilities in its communication with the Trusted Execution Environment, leading to a possible bypass of user consent, shown on Google's Pixel. Patching both leads to a provably correct protocol with intended APC user-consent guarantees, in the Universal Composability framework.

    Full text :  [ HAL ]  [ CISPA Link ]

  6. Return-oriented programming on RISC-V

    ASIA CCS 20, ACM Asia Conference on Computer and Communications Security, Taipei, 2020

    Prefix-code machine instructions allow hiding malicious instructions using unaligned jumps, crafting sequences of long (32-bit) instructions whose last 16 bits are either a valid instruction or a valid prefix that can be chained into overlapping sequences fooling ROP gadgets detectors. A tree-based detection method identifies them correctly.

    Full text :  [ ACM Link ]  [ ArXiv ]

See all Publications & Patents

Latest Positions

  1. Postdoc : Adversarial Training

    July 2025 - present, Dauphine University, Paris. LAMSADE / MILES team

    with Yann Chevaleyre (LAMSADE, Dauphine) and Rafaël Pinot (LPSM, Jussieu)

  2. PhD in Mathematics

    Oct 2021 - Jun 2025, INRIA - ENS, Paris. DYOGENE / ARGO Project-team

    Advised by Marc Lelarge and Kevin Scaman

    Construction and convergence of provably-correct neural networks.

Teaching

  1. Teaching Assistant : Deep Learning

    Deep Learning (MAP583) course by Kevin Scaman (INRIA - ENS), École Polytechnique

    Practical introduction to deep learning and all implementation details, with a focus on coverage of a large amount of different data domains and network architectures.

    Resources :  [ Synapses page ]  [ Practicals repository ]  [ Custom python package ]

  2. Guest Lecture : Neural network compression

    Deep Learning course by Marc Lelarge (INRIA - ENS), ENS Paris

    Introduction to neural network compression concepts and recent results, with a focus and practical session on activation reconstruction.

    Resources :  [ Lecture slides ]  [ Practical Session ]  [ Practical Session Solution ]

View Resume