What Does a Neuron Do? @ WBE Workshop 2023

In this talk, Mitya Chklovskii discusses the complexity of neurons and proposes a theoretical framework for understanding their function. He argues that the conventional view of neurons as feed-forward devices is oversimplified, and instead suggests that they should be considered as feedback controllers. This hypothesis eliminates the need for error backpropagation signals and allows neurons to learn from local information. Chklovskii introduces the direct data-driven control framework, which aligns with physiological observations and presents itself as a viable model for simulating neurons. He also discusses the challenge of inferring the degrees of freedom in neuronal dynamics and proposes learning the system’s parameters based on how well they can be learned. The speaker concludes by suggesting that neurons could implement a similar approach to adapt to unknown environments without needing a mechanistic model.