Thalamocortical circuits naturally perform computationally efficient hierarchical clustering.

Poster Session F - Tuesday, April 16, 2024, 8:00 – 10:00 am EDT, Sheraton Hall ABC

Charles Liu¹ (charles.liu.th@dartmouth.edu), Eli Bowen¹, Richard Granger¹; ¹Dartmouth

INTRODUCTION. Current artificial neural networks are based on a tiny number of salient neural characteristics (parallel operation of simple computing nodes), thus assuming that other characteristics (e.g., asymmetry of excitatory and inhibitory cells; different cell types (e.g., pyramidal, stellate, tonically active, modulatory); distinct circuit wiring patterns in different brain regions; etc.) may be computationally irrelevant. We investigate whether such additional neural characteristics may confer novel powerful algorithmic abilities. METHODS. We first simulated (as networks of neurons), and then abstracted into algorithm form, a thalamocorical circuit incorporating key features of reciprocal feedforward and feedback connectivity. Features include differential time courses of excitatory vs. inhibitory postsynaptic potentials, differential axon reach of pyramidal cells vs. interneurons, and different laminar afferent and projection patterns. RESULTS. The simulation was shown to organize stored memories into similarity-based categories (clusters), but also, topographically organized feedback then subtracted (inhibited) cluster information from inputs, such that subsequent feedforward operation produced successive subclusters. Analysis showed the relationship of these operations with well-studied algorithms for hierarchical clustering. Moreover, we demonstrated that the derived novel algorithms exhibited desirable computational space and time complexity, and corresponding scaling and efficiency characteristics. CONCLUSIONS. A novel algorithm for hierarchical clustering emerged from a relatively straightforward model of feedforward and feedback activity in a simulated thalamocortical circuit. The resulting algorithm compared well against the large literature of extant hierarchical clustering methods. Implications are discussed for further extraction and analysis of unexpected algorithms directly from brain circuit layout and operation.

Topic Area: PERCEPTION & ACTION: Other