The brain uses recurrent spiking neural networks for higher cognitive functions such as symbolic computations, particularly, processing and manipulation of sequences of symbols. While brains, seemingly without effort, process new information in the light of recent experiences and dynamic rules, models for neural networks of the brain have inherent difficulties in carrying out such computations. We show here that complex cognitive tasks that often contain a significant temporal processing component, including the requirement to flexibly incorporate task context and/or rule changes, can be solved by models of spiking networks if one includes neurons with spike frequency adaptation (SFA). We demonstrate the performance of a rather small spiking neural network (SNN) on the 12AX task which tests the ability to apply dynamic rules when detecting specific subsequences in a long sequence of symbols, and mathematical computations, such as comparison of two bit-strings encoding numbers, or evaluation of nested arithmetic expressions. These results suggest that tasks in which symbols are manipulated can be learned by models of spiking neural networks (SNNs) and even trained using a biologically plausible learning rule called e-prop. Hence, having SNNs that produce brain-like computational capabilities could enhance the convergence of further biological experiments, models, and theories for uncovering the computational primitives of the brain.
|Published - 22 Sept 2021
|Bernstein Conference 2021 - Online, -
Duration: 21 Sept 2021 → 23 Oct 2021
|Bernstein Conference 2021
|21/09/21 → 23/10/21