Timezone: »

 
Gradient Starvation: A Learning Proclivity in Neural Networks
Mohammad Pezeshki · Sékou-Oumar Kaba · Yoshua Bengio · Aaron Courville · Doina Precup · Guillaume Lajoie

@
in Over-parameterization: Pitfalls and Opportunities »

We identify and formalize a fundamental gradient descent phenomenon leading to 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 imbalances 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 but simple 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.

Author Information

Mohammad Pezeshki (Mila, Université de Montréal)
Sékou-Oumar Kaba (Mila, McGill University)
Yoshua Bengio (Mila - Quebec AI Institute)
Aaron Courville (University of Montreal)
Doina Precup (DeepMind)
Guillaume Lajoie (Mila, Université de Montréal)

More from the Same Authors