Developmental Evolution of Dendritic Morphology in a Multi-Compartmental Neuron Model - Artificial Neural Networks, 1999. ICANN 99. Ninth International Conference on (Conf. Publ. No. 470

Through the use of a multi-compartmental neuron simulation, Mainen and Sejnowski demonstrated that spike generation in neurons is a function of their dendritic structure [l]. In this paper we investigate the determination of dendritic morphology given a desired set of spike traces. A genetic algorithm is used to identify optimal parameters for a developmental model which simulates the growth of 3-dimensional dendrites. For two ’classes of ’neurons with different spiking behaviour, the developmental evolutionary process discovers ranges of viable dendritic morphologies which satisfactorally match the desired spike traces. 1 Compartmental Neuron Simulations The field of computational neuroscience demonstrates that by reducing the levels of complexity found in biology to computational models, biologically defensible results can be achieved [2]. Detailed compartmental models of neurons for example, can be accurately simulated using a number of software packages [3,4]. Using the NEURON simulation environment [3], Mainen and Sejnowski showed.that dendritic structure can determine the pattern of spike trains in neurons [l]. In Mainen and Sejnowski’s model a stimulus is injected into the soma of a neuron and is backpropagated through the neuron’s dendritic tree [1,5]. Due to the structure of the dendritic tree and the parameters of the compartmental model, spikes of neural activity are generated within the dendritic tree. The frequency and characteristics (single or mul~ tiple spikes) of spike generation were determined to be a function of the dendritic structure of the neurons. If dendritic structure is a key arbiter of firing patterns in neurons, we are interested in identifying dendritic structures given a desired spike train. Adapting neural signalling in compartmental neuron models has previously been studied using evolutionary computation and other optimisation techniques [6,7]. These experiments were performed on fixed neuron structures and changes in spike generation were the result of varying the parameters of the compartmental model. An alternative approach, adopted in this paper, is to keep the parameters of the compartmental model fixed and adapt dendritic morphology. We are interested in the design of complex, artificial neural systems using a combination of biologically-inspired algorithms from neural development and evolutionary computation. The motivation for the work in this paper is to assess the feasibility of evolving dynamic neural signalling in artificial neurons, beyond simple ‘integrate-andfire’ models. We aim to examine whether such functionality can be achieved simply by choosing arbitrary, artificial morphologies or whether there is a strong causal relationship between morphology and functionality. This paper thereby explores how spike trains in artificial neural systems can be generated through the adaption of dendritic structure. 2 Developmental Evolution We have implemented a 3D model of biological development, in which neuron-to-neuron connectivity is created through interactive self-organisation [8]. Development occurs as a number of overlapping stages, which govern how neurons extend axons and dendrites, collectively termed neurites. Neurons grow within an artificial, embryonic environment, into which neurons and their neurites emit local chemical gradients. The growth of neurites is influenced by the local gradients and the following sets of interacting, developmental rules: Intrinsic rules control the times at which neurites branch and the directions of growth Artificial Neural Networks, 7 10 September 1999, Conference Publication No. 470 0 IEE 1999 383