Separation Power of Equivariant Neural Networks
Two new research papers explore the separation power and universality of equivariant neural networks. The first paper characterizes inputs indistinguishable by such models and analyzes how hyperparameters like activation functions and depth influence their expressivity. It finds that non-polynomial activations are equivalent in expressivity and that depth improves separation power up to a certain point. The second paper establishes a universality theorem for invariant networks and introduces "entry-wise separability" for equivariant networks, demonstrating that depth and readout layers are crucial for achieving universality. AI
IMPACT These papers offer theoretical insights into the capabilities and limitations of specific neural network architectures, potentially guiding future model design.