The Strengths and Limitations of Equivariant Neural Networks
Robin Walters, Assistant Professor, Northeastern
Despite the success of deep learning, there remain challenges to progress including dataset size, generalization, and lack of guarantees. Incorporating symmetry into neural networks gives equivariant neural networks (ENN) which have helped address these challenges. I will discuss several dynamics applications, such as trajectory prediction, ocean currents, and robotics. However, there are also limits to the effectiveness of ENNs. In many applications where symmetry is only approximate or does apply across the entire input distribution, equivariance may hurt model performance. I will discuss recent work characterizing errors resulting from mismatched symmetry biases which can be used for model selection.
