Emergent Neural Network Mechanisms for Generalization to Objects in Novel Orientations

TMLR Paper4883 Authors

Abstract: The capability of Deep Neural Networks (DNNs) to recognize objects in orientations outside the training data distribution is not well understood. We investigate the limitations of DNNs’ generalization capacities by systematically inspecting DNNs' patterns of success and failure across out-of-distribution (OoD) orientations. We present evidence that DNNs (across architecture types, including convolutional neural networks and transformers) are capable of generalizing to objects in novel orientations, and we describe their generalization behaviors. Specifically, generalization strengthens when training the DNN with an increasing number of familiar objects, but only in orientations that involve 2D rotations of familiar orientations. We also hypothesize how this generalization behavior emerges from internal neural mechanisms – that neurons tuned to common features between familiar and unfamiliar objects enable out of distribution generalization – and present supporting data for this theory. The reproducibility of our findings across model architectures, as well as analogous prior studies on the brain, suggests that these orientation generalization behaviors, as well as the neural mechanisms that drive them, may be a feature of neural networks in general.

Submission Length: Long submission (more than 12 pages of main content)

Assigned Action Editor: ~Andrew_Kyle_Lampinen1

Submission Number: 4883