Invariant Odor Recognition with ON-OFF Neural Ensembles
Invariant recognition of a stimulus is a challenging pattern-recognition problem that must be dealt with by all sensory systems. Since neural responses evoked by a stimulus could be perturbed in a multitude of ways, could a single scheme be devised to achieve this computational capability? We examined this issue in locust olfactory system. We found that odor-evoked responses in individual projection neurons in the locust antennal lobe varied unpredictably with repetition, stimulus dynamics, stimulus history, presence of background odorants, and changes in ambient conditions. Yet, a highly-constrained Bayesian logistic regression approach with ternary weights could provide robust odor recognition. We found that this approach could be further simplified: sum firing rates of ON neurons and subtract total activity in OFF neurons ( ON minus OFF classifier). Notably, we found that this approach could be generalized to develop a Boolean neural network that can perform well in a non-olfactory pattern recognition task.