The olfactory system is a remarkable system capable of
discriminating between very similar odorant mixtures. It is thought that this
function is somewhat related with spatiotemporal activity patterns generated in
mitral and tufted (M/T) cells of the olfactory bulb (OB) during odor presentation
(1). These spatiotemporal patterns are in part controlled by reciprocal
dendrodendritic synapses between M/T cells and granule cells (GCs). The spatial
aspect of these patterns would be provided by contrast enhancement mechanisms
generated by GC-mediated lateral inhibition, whereas the temporal aspect would
result from synchronizing and desynchronizing processes conveyed by reciprocal
feedback interactions between M/T cells and GCs (2). The efficient coding
scheme for high-dimensional patterns used by the OB could be explored by
artificial pattern recognition systems. Inspired by this, we developed a
spiking neural network OB model and present here test results from its
application to recognition tasks involving patterns taken from both artificial
and real databases. The model consists of a simplified version of OB containing
only mitral cells (MCs) and GCs. These two cell types were modeled according to
the Izhikevich formalism (3). The two cell populations were arranged into two
square arrays with the same size. The GC array has four times the number of
cells in the MC array. The number of cells in the MC array depends on the
dimension of the input space. The probability of a synaptic contact between a
MC and a GC was given by
P(
d) = exp[−(
d/
λ)
2], where
d is the horizontal distance between them and
λ is a parameter. Signal propagation times were neglected. For each
established contact between MC and GC grids two synapses were created: one
excitatory from the MC to the GC, and one inhibitory from the GC to the MC.
These synapses were modeled as alpha functions with short-term facilitation and
depression as in the dynamic synapse model (4). The output of the system is a
vector with the same dimension as the number of cells in the MC array. The
components of this vector are the MCs firing rates. The system thus transforms
input patterns into spatiotemporal output patterns determined by its intrinsic connectivity
and synaptic parameters. These dynamic output patterns can be transformed into
stable states using readout modules (5). Virtually any type of input data can
be fed into the system, perhaps requiring some pre-processing. We tested our
system with five datasets: two were artificially created by us and are composed
of spatial patterns with different overlap degrees, and three were taken from
the UCI repository (6). These latter were “BC” (breast cancer), “Heart” and “Parkinsons”.
These three datasets have instances with missing attributes – especially Heart
–, which makes the classification task hard. All input patterns were converted
to spike trains using a noisy integrate-and-fire model. The performance of the
system was evaluated with respect to two parameters:
λ and a synaptic weight parameter. Results for the artificial patterns
show that there is an optimal range of parameters for which the system
discriminates well patterns with high overlap degree. For strong synaptic
weights the distance measure used to discriminate between patterns oscillates
in time so that a threshold detector can be used as readout. Good
classification performances for the three real-world datasets were obtained
with the use of a parallel perceptron as readout. Without fine-tuning of the
Parallel Perceptron parameters, classification rates for the three datasets
were above 75% for BC and Heart and above 70% for Parkinsons.
Support: ACR is supported by a CNPq research grant.
Work also supported by INCeMaq.
References
1.
Lledo, P.M., Gheusi, G., and Vincent,
J.D. (2005). Information processing in the mammalian olfactory system. Physiol. Rev. 85, 281-317.
2.
Cleland, T.A., and Linster, C. (2005). Computation in
the olfactory system. Chem. Senses
30, 801-813.
3.
Izhikevich, E.M. (2007) Dynamical
Systems in Neuroscience: the geometry of excitability and bursting. Cambridge,
MA: MIT Press.
4.
Tsodyks, M., Pawelzik, K., and Markran,
H. (1998). Neural networks with dynamic synapses. Neural Comput. 10, 821-835.
5.
Maas, W., Natschläger, T., and Markran,
H. (2002). Real-time computing with stable states: a new framework for neural
computation based on perturbations. Neural
Comput. 14, 2531-2560.
6.
Frank, A., and Asuncion, A. (2010). UCI
Machine Learning Repository. http://archive.ics.uci.edu/ml. Irvine,
CA: University of California, School of Information and Computer Science.