Contributed Talk Sessions | Poster Sessions | All Posters | Search Papers

Poster Session C: Friday, August 15, 2:00 – 5:00 pm, de Brug & E‑Hall

What governs the emergence of brain-like specialized neurons in artificial neural networks?

Brian S Robinson¹, Michael Bonner²; ¹Johns Hopkins University Applied Physics Laboratory, ²Johns Hopkins University

Presenter: Brian S Robinson

Neurons with specialized properties have been widely characterized across the brain. It is a broad question as to why this occurs, with computational theories often assuming a central role of nonlinear neuron activation functions and connectivity constraints. In this work, we characterize neuron-specialization in artificial neural networks by extending regression-based approaches for predicting experimentally recorded neural activation patterns. When investigating a range of performant artificial neural network architectures, we demonstrate that (1) brain-aligned specialized neurons can emerge in layers without nonlinear neuron activation functions, and (2) the emergence of brain-aligned specialized neurons depends on training properties, not strictly on architecture. Overall, this work suggests that new and complementary explanations for the emergence of specialized neurons in biological brains may be needed, such as processes underlying learning and optimization. Furthermore, this work motivates brain-to-model comparison techniques that respect and further investigate properties of neuron specialization. These results may additionally inform general interpretability approaches for artificial neural networks, where methods for obtaining units for inspection is an active area of research.

Topic Area: Brain Networks & Neural Dynamics

Extended Abstract: Full Text PDF