Skip to content Skip to navigation

A Biologically Constrained Spiking Neural-Astrocytic Network Learns Action Habits for Autonomous Mobile Robots

Principal Investigator: 
Grant Agency: 
Grant Duration: 
06/15/2018 to 06/14/2020
[Goal] To devise an end-to-end brain-morphic algorithm that 1) is constrained by the neural connectome associated with the targeted function; 2) incorporates astrocytes, the most abundant brain cells, with newly discovered computational roles; and 3) controls a mobile robot that autonomously learns to exhibit action habits, by associating sensory and location information with its behavior. 
[Scopes] We will 1) translate into Intel's Loihi our biologically-constrained SNN to endow mobile robots with spike-based active simultaneous localization, mapping (SLAM) and planning; 2) introduce astrocytes into Loihi’s x86 cores and grow Spiking Neural-Astrocytic Networks (SNANs), to enable astrocytic learning of action sequences for a given neural representation of a stimulus; 3) Benchmark our robots in developing habits of quick deciding on their path finding in a real environment. 
[Impact] By demonstrating for the first time how robotic functions can emerge naturally from their controller’s structure, our brain-derived SNAN architecture will eliminate the need for assuming all-to-all initial connectivity and maximize training efficiency, since a small number of synaptic connections suffices to map an environment. Further, introducing the long-neglected astrocytes into practical brain-morphic applications will 1) extend learning beyond changes in the strength of hardwired neuronal synapses and onto the processes that astrocytes have to modulate neurons; as well as 2) introduce an orthogonal computational dimension to neuronal processing, with distinct dynamics and temporal scales that resemble those of human behavior. 
Overall, the proposal aims to 1) leverage the mounting knowledge on the brain connectome; and 2) introduce astrocytic learning, with its own, slow, dynamics, resembling those of behavioral responses. Having a pre-defined architecture, the training of our spiking neural networks (SNNs) is much more efficient, especially when compared to deep neural networks that need thousands of labeled data to trim their initial all-to-all connectivity into a structure serving the targeted function. By relying on astrocytes to bridge the temporal gap between fast neural and slow behavioral activities, our SNANs are directly applicable to real-time control of a robotic behavior in general, and movement in particular.