On the Relation between Gradient Directions and Systematic Generalization

Yuanpeng Li, Liang Zhao, Joel Hestness, Kenneth Church, Mohamed Elhoseiny

Keywords: Systematic generalization, Compositional generalization

Abstract: Systematic generalization is a critical property that most general deep learning algorithms lack. In this paper, we investigate the relation between gradient directions and systematic generalization. We propose a formulation to treat reducible training loss as a resource, and the training process consumes it to reduce test loss. We derive a bias that a training gradient is less efficient in using the resource at each step than an alternative gradient that leads to systematic generalization. The bias is avoided if and only if both gradients are zero or point in the same direction. We demonstrate the bias in standard deep learning models, including fully connected, convolutional, residual networks, LSTMs, and (Vision) Transformers. We also discuss a requirement for the generalization. We hope this study provides novel insights for improving systematic generalization.

Primary Area: causal reasoning

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors' identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 7135