A Push-pull Model of Orientation-selectivity in a Hybrid Vlsi System of Spiking Neurons

We describe the use of a programmable, rewireable, multi-chip aVLSI neu-ronal system in creating orientation-selective spiking neurons with various receptive fields using a feedforward model. The multi-chip system consists of a silicon retina, an asynchronous router board, and a transceiver multi-neuron chip. The three modules communicate using the asynchronous ar-bitered address-event representation (AER) protocol. This protocol uses asynchronous digital pulses, similar to spikes in a neuronal system. The circuit on the multi-neuron chip approximates a cortical microcircuit. The receptive fields of the neurons are configured by the virtual connections of selected sets of pixels on the silicon retina. These connections are effected through the programmed look-up table on the field programmable gate-array (FPGA) chip on the router board: This table sets the destination addresses for the incoming retina addresses. We used the multi-chip spike-based system to synthesize orientation-tuned cortical-like neurons with different receptive fields. In particular, we used this system to explore the neuron's response for two feedfor-ward models. In the first model, the retinal inputs drive the neuron only through excitatory synapses, and in the second model, the retinal inputs drive the neuron through both excitatory and inhibitory synapses similar to a push-pull model [Hubel and Wiesel, 1962]. We found that a neuron whose orientation-selectivity comes from a push-pull model arrangement of the retinal inputs is more robust to the variation in the timing of the input spikes and responds over a wider range of stimulus speeds than a neuron whose receptive field is generated purely through an excitatory feedfor-ward model. Network of neurons Orientation-selective neurons Figure 1: Block diagram of a multi-chip system in which virtual connections from a set of neurons on a silicon retina onto another set of neurons on a transceiver chip are described using the router module. The router communicates with the multi-neuron chip if the spike address from the retina falls within the lookup table programmed on this module.