Cortical processing on distributed networks of silicon spiking neurons
Cortical neural networks are characterized by a large degree of recurrent excitatory connectivity and local inhibition. This type of connectivity among neurons is remarkably similar across all areas in the cortex. It has been argued that a good candidate model for a canonical micro-circuit, potentially used as a general purpose cortical computational unit in the cortices, is the soft Winner-Take-All (WTA) circuit. A neural circuit performing soft WTA can be modeled using spiking neurons with local excitatory and global inhibitory couplings. These models can be efficiently implemented in silicon using Integrate & Fire neurons and dynamic synapses.
We developed a set of VLSI devices comprising arrays of silicon neurons and dynamic synapses, with recurrent excitatory and global inhibitory couplings, to implement soft WTA behaviour. Using the Address-Event-Representation (AER), our chips can be arbitrarily coupled together to form large multi-chip networks.
We present a distributed multi-chip neural network consisting of two multi-neuron chips exhibiting soft WTA behaviours, coupled to silicon vision sensor, to demonstrate cortical-like signal processing with real visual inputs.