The Efficient Computing Hypothesis Normative Theories of Neural Computation

The defining question of systems neuroscience is `how does neural activity relate to computation?’ Normative theories are one of many useful approaches to answering this question. They proceed by modelling neural activity as the solution to an optimisation problem. The paradigmatic example is the efficient coding hypothesis, in which neural activity is modelled as a faithful but efficient encoding of information. Efficient coding theory is widely successful, but mainly in sensory systems whose purpose seems well approximated as sending compact summaries of information to cortex. We argue that the rest of the brain is not just communicating information, it is computing with it, and our normative theories should reflect that. The other major normative approach, connectionism, aligns with this idea. It models neural activity using a task-optimised artificial neural network, and since these networks can perform most tasks (think of ChatGPT), connectionism can be used to study almost any neural computation. However, the complexity of these models sacrifices tractability: neurons in artificial networks are often as inscrutable as their biological cousins. Therefore, this thesis introduces a tractable normative framework for studying the neural implementation of computations, that we term the efficient computing hypothesis. This framework takes the efficient coding optimisation problem, and simply adds a constraint that ensure the predicted neural activity can implement the relevant computation. Most of this thesis is devoted to using this theory to normatively model two cortical representations - path-integrating representations like the entorhinal grid code, and structured working memory representations in prefrontal cortex. A minor theme in this thesis is using similar ideas to study the modularity of neural circuits. In sum, by framing a tractable normative theory for the neural implementation of computations, we take small steps towards understanding and inferring the brain’s algorithms from neural activity, and predicting its puzzling neural implementations.