Spatially-embedded Recurrent Neural Networks: Bridging common structural and functional findings in neuroscience, including small-worldness, functional clustering in space and mixed selectivity
Jascha Achterberg, Danyal Akarca, Duncan Astle, John Duncan, University of Cambridge, United Kingdom; Daniel Strouse, Matthew Botvinick, DeepMind, United Kingdom
Posters 2 Poster
Pacific Ballroom H-O
Fri, 26 Aug, 19:30 - 21:30 Pacific Time (UTC -8)
Brain networks exist in a biophysical world. In this world, the brain as a whole needs to control and navigate its organism within a complex environment, while its constituent neurons must economically balance the resources they use to grow connections. To build and sustain connections, they need to overcome the strain caused by long distances in physical 3D and their own topological space. Due to being exposed to the same basic forces and hence optimization problem, many brains converge on similar features in their structure and function. To observe the effects of these basic forces on a network’s optimization process, we introduce the spatially-embedded RNN (seRNN). We find that seRNNs, due to existing in 3D Euclidean and topological space, naturally converge on solving a task using modular small-world networks in which functionally similar units cluster in space and utilise a mixed selective code. This shows (a) how fundamental, but seemingly unrelated, neuroscientific findings can be attributed to a network’s biophysical optimization process and (b) how spatially-embedded neural networks can serve as model systems to bridge between structural and functional research communities to move neuroscientific understanding forward.