Inter-animal transforms as a guide to model-brain comparison

To address the question of how to compare DNN model activations to brain data, we investigate what transforms best describe similarity in neural activity in the same brain area between conspecifics. We expect neural responses to be functionally highly similar within a species (since we expect findings to generalize across animals). What kind of transform will make such similarity most evident? That is, under what kind of transform are conspecifics’ neural responses highly similar to each other? Researchers often default to linear regression as a reasonable transform class for measuring neural response similarity. We propose an improved transform class that uses a generalized linear model (GLM) whose noise matches the approximately Poisson noise in the neural data, and whose non-linear link function is akin to the activation function of a biological neuron. Incorporating these biologically motivated constraints into the inter-animal transform class substantially improves similarity scores compared to linear regression. We then build a DNN model of mouse visual cortex that swaps out ReLU activations for a more biologically plausible softplus activation function, combined with Poisson noise, to produce activations that are more similar to neural responses. We find that a Poisson GLM whose link function exactly matches the model activation function again yields the highest similarity scores between different randomly seeded instances of our softplus models. This result gives mechanistic insight into why the best performing animal transform class has a non-linear link as well as Poisson noise structure. Moreover, we show that our Poisson GLM not only achieves higher similarity scores for the same layer between model instances, but also scores activations in model layers that are physically far apart as highly dissimilar to each other. Finally, we estimate the number of neurons and number of stimuli that would need to be recorded to accurately estimate inter-animal similarity.