We describe a procedure for explaining neurons in deep representations by
identifying compositional logical concepts that closely approximate neuron
behavior. Compared to prior work that uses atomic labels as explanations,
analyzing neurons compositionally allows us to more precisely a