UNIVERSITY OF HERTFORDSHIRE COMPUTER SCIENCE RESEARCH COLLOQUIUM presents "An Artificial Life Perspective on Olfactory Systems: Evolving Neural Coding, Developmental Symmetry and Odour Recognition in Agents" Nicolas Oros (School of Computer Science, University of Hertfordshire, UK) 10 March 2010 (Wednesday) Lecture Theatre E351 Hatfield, College Lane Campus 3 - 4 pm Everyone is Welcome to Attend Refreshments will be available Abstract: I will present the research conducted during my PhD. My thesis addressed the problem of creating simulated agents controlled by neural networks that share features with biological olfactory systems. This work draws from the fields of Artificial Life, Artificial Intelligence and Neuroscience. My main research questions were: * What computational strategies does a network of spiking neurons use to discriminate odours and control an agent? * How such a neural architecture can encode information (as spatio-temporal patterns using different neural coding strategies) ? * How to evolve a neural controller for agents that can encode temporal delays using spiking neurons and a developmental approach? * How can a spiking neural network encode information in order to control an agent that is attracted by a low level of concentration but repelled by a high level of the same chemical concentration? Is it necessary to have different types of sensory neuron that react to different concentration values in order to perform this task? The techniques used in this work included simulated agents and chemicals situated in a 2D environment, spiking neural controllers in which neurons were placed on a 2D substrate and transmission delays depended on the length of the connections, a developmental model used with an indirect encoding that could map a genome onto a neural network, and a genetic algorithm used to evolve controllers. The findings of this program raised several interesting issues. Results have shown that using a biologically plausible sigmoid function in my model to map chemical concentration to the total input current of a leaky integrate-and-fire neuron, agents were able to detect the whole range of chemical concentration as well as small variations. The sensory neurons used in this work are able to encode the stimulus intensity into appropriate firing rates. This research also reveals that two different neural coding strategies can be used by a simple neural network to control an agent . Both temporal coincidence (of spikes) and firing rate encoding strategies were important mechanisms used by the same neural network in different environmental conditions. In addition, realistic model of neural noise were shown to improve the behaviour of an agent to perform a task like chemotaxis. Models used to evolve developmental neural controllers for agents have been created and results have shown that evolved agents could perform a relatively realistic and difficult task, and their neural controllers could encode information in space and time. In this work, the use of symmetrical structures was shown to have major benefits for the evolution of neural controllers. Finally, a detailed analysis of the neural dynamics was conducted on an evolved neural network and has shown that the model generates controllers that use rather sophisticated neural coding strategies involving detailed temporal information. This analysis revealed that single spikes sent at specific moments could modify the whole activity of a network and the behaviour of an agent. (As some people have seen my previous talk, I will try to spend most of the time on these last (NEW) results.) --------------------------------------------------- Hertfordshire Computer Science Research Colloquium http://homepages.stca.herts.ac.uk/~nehaniv/colloq