R
Rob Fergus
Researcher at New York University
Publications - 175
Citations - 103027
Rob Fergus is an academic researcher from New York University. The author has contributed to research in topics: Object (computer science) & Reinforcement learning. The author has an hindex of 82, co-authored 165 publications receiving 85690 citations. Previous affiliations of Rob Fergus include California Institute of Technology & University of Oxford.
Papers
More filters
Perturbing BatchNorm and Only BatchNorm Benefits Sharpness-Aware Minimization
TL;DR: This paper investigated the connection between two popular methods commonly used in training deep neural networks: Sharpness-Aware Minimization (SAM) and Batch 2 Normalization and found that perturbing only the affine BatchNorm parameters in the adversarial step of SAM benefits the generalization performance, while excluding them can decrease the performance strongly.
Proceedings Article
Automatic Data Augmentation for Generalization in Reinforcement Learning
TL;DR: This paper proposed three approaches for automatically finding an effective augmentation for any RL task and combined with two novel regularization terms for the policy and value function, required to make the use of data augmentation theoretically sound for actor-critic algorithms.
Posted Content
Decoupling Value and Policy for Generalization in Reinforcement Learning
Roberta Raileanu,Rob Fergus +1 more
TL;DR: In this article, the authors propose an auxiliary loss which encourages the representation to be invariant to task-irrelevant properties of the environment, which achieves state-of-the-art performance on the Procgen benchmark and outperforms popular methods on DeepMind Control tasks with distractors.
Low-level image priors and laplacian preconditioners for applications in computer graphics and computational photography
Rob Fergus,Dilip Krishnan +1 more
TL;DR: Novel image priors and efficient algorithms for image denoising and deconvolution applications, and effective preconditioners for Laplacian matrices for discrete Poisson equations are developed.
Journal Article
Empirically Verifying Hypotheses Using Reinforcement Learning
TL;DR: This paper formulates hypothesis verification as an RL problem, and builds an agent that, given a hypothesis about the dynamics of the world, can take actions to generate observations which can help predict whether the hypothesis is true or false.