Gradient Starvation: A Learning Proclivity in Neural Networks

By Mohammad Pezeshki, Sekouba Kaba, Yoshua Bengio, Aaron C. Courville, Doina Precup, Guillaume Lajoie. 2020

353 citations recorded by the source corpus.

We identify and formalize a fundamental gradient descent phenomenon resulting in a learning proclivity in over-parameterized neural networks. Gradient Starvation arises when cross-entropy loss is minimized by capturing only a subset of features relevant for the task, despite the presence of other predictive features that fail to be discovered. This work provides a theoretical explanation for the emergence of such feature imbalance in neural networks. Using tools from Dynamical Systems theory, we identify simple properties of learning dynamics during gradient descent that lead to this imbalance, and prove that such a situation can be expected given certain statistical structure in training data. Based on our proposed formalism, we develop guarantees for a novel regularization method aimed at decoupling feature learning dynamics, improving accuracy and robustness in cases hindered by gradient starvation. We illustrate our findings with simple and real-world out-of-distribution (OOD) generalization experiments.

Subfields: Artificial Intelligence, Statistical and Nonlinear Physics

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Artificial Intelligence

In Statistical and Nonlinear Physics

Citations in the corpus

Continual Learning via Local Module Composition (2021, methodology): “Note, that model selection is an unsolved challenge in the OOD generalization literature, where selecting model using oracle strategies is sometimes excepted if the baselines methods are also tuned using oracle strategies [30, 77].”
Just Train Twice: Improving Group Robustness without Training Group Information (2021, background): “Pezeshki et al. (2020) modify the dynamics of stochastic gradient descent to avoid learning spurious correlations.”
Self-Supervised Learning with Data Augmentations Provably Isolates Content from Style (2021, background): “[96] Mohammad Pezeshki, Sékou-Oumar Kaba, Yoshua Bengio, Aaron Courville, Doina Precup, and Guillaume Lajoie.”
Last Layer Re-Training is Sufficient for Robustness to Spurious Correlations (2022, methodology): “Finally, we can leverage better representation learning methods [76], including self-supervised learning methods [e.”
Objective Robustness in Deep Reinforcement Learning (2021, background): “• The true reward R is sparse, but the proxy R′ is dense [31].”
Invariance Principle Meets Information Bottleneck for Out-of-Distribution Generalization (2021, background): “, 2019) and several other works were proposed (Ahuja et al., 2020; Pezeshki et al., 2020; Krueger et al., 2020; Robey et al., 2021; Zhang et al., 2021).”