Bispectral Neural Networks

TL;DR

Bispectral Neural Networks (BNNs) are designed to learn invariant and complete representations of signals under group actions, enabling robust and symmetry-aware learning from data.

Contribution

This paper introduces BNNs that leverage the bispectrum to achieve invariant, complete representations and learn symmetries directly from data.

Findings

BNNs can learn group symmetries and irreducible representations.

BNNs exhibit strong invariance-based adversarial robustness.

BNNs effectively preserve signal structure while removing group-induced variation.

Abstract

We present a neural network architecture, Bispectral Neural Networks (BNNs) for learning representations that are invariant to the actions of compact commutative groups on the space over which a signal is defined. The model incorporates the ansatz of the bispectrum, an analytically defined group invariant that is complete -- that is, it preserves all signal structure while removing only the variation due to group actions. Here, we demonstrate that BNNs are able to simultaneously learn groups, their irreducible representations, and corresponding equivariant and complete-invariant maps purely from the symmetries implicit in data. Further, we demonstrate that the completeness property endows these networks with strong invariance-based adversarial robustness. This work establishes Bispectral Neural Networks as a powerful computational primitive for robust invariant representation learning