Peter Jurica, RIKEN BSI, Japan

Sergei Gepshtein, RIKEN BSI, Japan

Ivan Tyukin, RIKEN BSI, Japan; University of Leicester, UK

Danil Prokhorov, Toyota Technical Center, Ann Arbor, USA

Cees van Leeuwen, RIKEN BSI, Japan

We propose a design for adaptive optimization of sensory systems. We consider a network of sensors that measure stimulus parameters as well as the uncertainties associated with these measurements. No prior assumptions about the stimulation and measurement uncertainties are built into the system, and properties of stimulation are allowed to vary with time. We present two approaches: one is based on estimation of the local gradient of uncertainty, and the other on random adjustment of cell tuning. Either approach steers the network towards its optimal state.

Illustration of evolution by the random search algorithm driven by implicit uncertainties in the local optimization framework. This evolution proceeds slower than evolution by gradient descent, which is why we show the states of the system separated by a greater number of iterations than in the previous figures. In the optimized network, high sensitivity tends to concentrate along the local optimal set, but the random search algorithm forces the cells to continuously change their tuning parameters, which is why the distribution of sensitivity in a network optimized by random search is more dispersed than in a network optimized by gradient descent. 

Publications:

Jurica, P., Gepshtein, S., Tyukin, I., Prokhorov, D. & van Leeuwen, C. (in press). Unsupervised adaptive optimization of motion-sensitive systems guided by measurement uncertainty. Peer-reviewed article  to appear in Proceedings of The Third International Conference on Intelligent Sensors, Sensor Networks and Information Processing 2007 (ISSNIP 2007).