Hardware architecture of a neural network model simulating pattern recognition by the olfactory bulb

We designed, built, and tested an electronic neural network which replicates many features exhibited by the olfactory bulb. The nonlinear dynamics of the network are motivated by experimental findings in EEG recordings of the bulb, which indicate that a massively parallel architecture can be used to best describe the bulb's dynamics. The electronic design is a digitallanalog hybrid approach, utilizing the speed and flexibility of random access memory (RAM) for the initial storing and further modification of synaptic strengths, while still preserving the analog computational power of neural networks. A simple “learning” algorithm is implemented to show qualitative agreement to experimental results. The hardware design also includes a multiplexing scheme which decreases the number of connections, in the hardware, from order N2 to order N, where N is the number of neural units. The input is a parallel set of high and low step functions. The key operation is a bias that is induced by a step input, causing the system to switch from a low-gain equilibrium state to a high-gain limit cycle state. The output is read as a spatial pattern of oscillatory amplitudes. These results support our physiological explanations of how the olfactory system categorizes odors.