Training a neural model using the C. elegans connectome to perform exploration tasks

Society for Neuroscience 2019


Abstract

Understanding the link between information processing in neural circuits and behavior remains a key goal in neuroscience. Network circuits extracted from the brain can be represented as a graph, where neurons are nodes and synapses are edges; attributes such as edge weights can provide additional context and information about the flow of information through the network. With the addition of dynamic neuronal models, these connectomes can be modeled as a system that takes sensory information as inputs and produces outputs that act on the surrounding environment, process sensory input and produce output. Simulating this dynamic model can help interpret experimental results and aid hypothesis development. Simulation can be done using a variety of tools with various levels of fidelity and objective functions (e.g., performance, low-level biological fidelity, concurrence with neuroscience theory). Given the extreme simplicity of the C. elegans nematode and its complete connectome, it is an optimal candidate for initial research discovery. Using simple neuron models, its entire nervous system can be readily simulated and we have demonstrated this on low-level hardware. In a model of C. elegans running in a python simulation environment, we investigate the exploration behavior of C. elegans by modifying the weights of a C. elegans connectome with a genetic algorithm. Therefore, we have investigated training the connectome to perform simple tasks in simulation by using a simple integrate-and-fire neuron model and a simple kinematic model of an agent. Example environments include gradient following, finding food, avoiding collisions, and noxious stimuli. We study our trained connectome in changing environmental conditions and investigate the effects of ablating neurotransmitter pathways on behavior. This work has implications for low-complexity, bio-inspired robotic exploration algorithms which may be more robust than reinforcement learning methods using artificial neural networks.

Citation

Contact Us


Chief
Ashley Llorens
Ashley.Llorens@jhuapl.edu
240-228-0312

Physical Address
7701 Montpelier Road
Laurel, MD 20723


The Intelligent Systems Center is located at the Montpelier Campus of the Johns Hopkins Applied Physics Laboratory.
Click here for a map, directions and other visitor information.