Neural Processing of Counting in Evolved Spiking and McCulloch-Pitts Agents

Abstract

This paper investigates the evolution of autonomous agents that solve a memorydependent counting task. Two types of neurocontrollers are evolved: networks of McCulloch-Pitts neurons, and spiking Integrate-And-Fire networks. The results demonstrate the superiority of the spiky model in terms of evolutionary success and network simplicity. The combination of spiking dynamics with incremental evolution leads to the successful evolution of agents counting over very long periods. Analysis of the evolved networks unravels the counting mechanism and demonstrates how the spiking dynamics are utilized. Using new measures of spikiness we find that even in agents with spiking dynamics, these are usually truly utilized only when they are really needed, i.e., in the evolved subnetwork responsible for counting.

Axiomatic Scalable Neurocontroller Analysis via the Shapley Value

“…This was achieved by defining a different function for each characteristic of interest and running the MSA on the multi-perturbation data gathered for each function. Saggie et al [25] focus on the study of counting agents such as W10 mentioned in this paper. A careful definition of the function to be analyzed allowed the MSA to reveal those synapses specifically in charge of counting, differentiating them from the synapses which contribute to other aspects of the agent's performance.…”

“…It was shown that different perturbation magnitudes tend to reveal different aspects of the neural functioning, with the smaller magnitudes accentuating neural elements involved in longer range dynamics. Lastly, Saggie et al [25] evolved neurocontrollers composed of discrete-time integrate-and-fire neurons. In order to determine the contributions of the neurons' integration capability to the agent's overall performance (as opposed to the importance of the neuron's information content, as revealed by using stochastic lesioning), they used a perturbation method which simply set a perturbed neuron's membrane time-constant to zero, thereby rendering it a standard McCulloch-Pitts neuron.…”

Evolving Spiking Networks for Turbulence-Tolerant Quadrotor Control

“…As membrane potential provides implicit memory, spiking networks are able to produce temporally dynamic activation patterns (Maass, 1997). Saggie-Wexler et al (2006) highlight that this property makes them particularly suited for temporal problems such as robot control as steady-state errors and unmodelled dynamics can be accounted for. Recent studies in cognitive science highlight the importance of dynamic processes in memory and learning (Buzsaki, 2006).…”

Neurocontroller Analysis via Evolutionary Network Minimization

“…In order to compare the neural activity patterns of the original agents with those of the minimized agents, we evolved agents to solve a more complex task, based on the work presented in [14]. In this variant of the original task, in order for the agent to successfully eat, it must wait on the food item for 5 steps without moving or turning.…”