Evolving adaptive coincidence-detecting neurons

Meeting Details

For more information about this meeting, contact Carina Curto, Timothy Reluga.

Speaker: Garrett Mitchener, College of Charleston
(Host: Andrew Belmonte)

Abstract: I will describe a computational experiment in which a selection-mutation process evolves neuron-like cells, combining evolutionary and biochemical dynamics. The simulated organisms, called agents, are designed to resemble single cells, each of which has an internal state consisting of counts of abstract molecules, plus a genome that specifies how they interact. The goal of this project is to start with random genomes and subject them to selective breeding, mutation, and recombination so that they evolve the ability to detect coincidences in a spike train, an essential ability of living neurons. When two input spikes arrive separated by a short delay, the agent should fire an output spike of its own, but when spikes arrive widely separated, the agent should produce no output spike. Agents are then given an additional Hebbian learning task. After receiving many closely spaced spikes, they should fire more eagerly even when spikes arrive somewhat separated. After a period of low activity, they should fire more skeptically, only after spikes arrive very close together. The simulation generally succeeds, discovering genomes encoding reaction networks that transfer activity from input to output, but with feedback loops that inhibit the transfer and only allow it to succeed when input spikes are close. Some of these inhibitory reactions are themselves inhibited by sustained input activity, accomplishing the Hebbian learning task using a mechanism similar to that of NMDA receptors.

---

Room Reservation Information

Room Number: 106 McAllister

Date: 04/14/2015

Time: 1:00pm - 2:00pm