Divisive normalization as a mechanism for hierarchical causal inference in motion perception

Causal inference (CI) has been proposed as a universal computational motif in the brain [Shams & Beierholm 2020]. However, its neural implementation is unclear. Likewise, Divisive Normalization (DN) has been proposed as a canonical circuit motif [Heeger 2011], but there are competing theories about the computations that DN implements (gain control, attention). In this work, we unified both by showing how an extended DN model can account for neural predictions made by a CI model [Shivkumar et al. 2022] in the context of hierarchical motion perception.
Specifically, we generated CI predictions for neural responses to a center-surround stimulus. CI makes interesting predictions for neural responses encoding two latent variables in the model: retinal and relative velocities. These predictions (supported by preliminary data) resemble the responses of two previously described classes of neurons in area MT: those with antagonistic or integrative surround [Born & Bradley 2005]. We investigated whether DN could be a potential mechanism that implements CI computations by fitting it to the CI predictions. Classic DN postulates that each neuron’s response is divisively modulated by pooling the activity of neighboring neurons. We analytically showed that a normalization pool that only incorporates the activities of neurons responding to center and surround stimuli alone cannot explain the complex CI responses. Instead, we found that a normalization pool that additionally includes multiplicative interactions between the center and surround activities can explain our CI predictions. We used tuning properties of MT neurons [DeAngelis & Uka 2003] to generate both: (a) biologically-realistic CI predictions and (b) a realistic normalization pool. We showed that the same DN model can explain CI neural predictions across different center-surround speeds. Our results suggest that an extended DN architecture, with interaction terms in the normalization pool, may serve as a mechanism to implement CI at the neural circuit level.