We present a new image classification model that extends CNNs with biologically-inspired higher-order convolutions. Outperforms standard CNNs on benchmarks and shows unique representational properties, bridging neuroscience and deep learning.

We propose a stimulus-wise decomposition of the mutual information that is (1) principled (axiomatic justification), (2) tractable (estimated via diffusion models), and then we demonstrate it's application on a model of visual neurons.