Models of olfactory information processing on modern parallel hardware
CBN (Computational Biology and Neurocomputing) seminars
Friday 26 March 2010
to 11:00 at
Thomas Nowotny (Informatics, University of Sussex)
In this talk I will present our work on models of olfactory information processing in insects and their implementation on modern, parallel hardware. The insect olfactory system is a predominantly feedworward system of a distinct convergent - divergent - convergent structure. We have demonstrated in a series of papers that this, at first sight unintuitive, structure may be the biological correlate of what is commonly referred to as kernel methods in the machine learning community. The corresponding models of the insect olfactory pathway are feed-forward neuronal networks with 10s of thousands of neurons and 100s of thousands connections, which need considerable computational resources, in particular when considering larger timescales as in learning. After introducing the models I will discuss how the recently introduced CUDA framework, a general purpose computing application programmming interface for NVidia graphical processing units (GPUs), allows large speedups in the simulations. We have observed 7-11 fold speedups in connectionist models and an overall 18 fold speedup for a spiking neuronal network model when comparing performance on an NVidia Tesla S1070 device to a recent 3GHz AMD Phenom system. The talk will be accesible to a general audience interested in computational neuroscience and does not require in-depth knowledge in either olfaction or parallel computing.