Hybrid analog-digital architectures for neuromorphic systems

Signal restoration is necessary to perform computations of significant complexity. In digital computers each state variable is restored to a binary value, but this strategy is incompatible with analog computation. Nevertheless, cortical neurons, whose major mode of operation is analog, are able to perform prodigious feats of computation. The authors´ research on visual cortex suggests that cortical neurons are able to compute reliably because they are organized into populations in which the signal at each neuron is restored to an appropriate analog value by a collective strategy. The strategy depends on feedback amplification that restores an input signal towards a stored analog memory. This principle is similar to recall by autoassociative neural networks. Networks of cortical amplifiers can solve simple visual processing tasks. They are well-suited to sensory processing because the same principle that restores their analog signals can also extract meaningful features from ambiguous sensory input. The authors describe a hybrid analog-digital CMOS architecture for constructing networks of cortical amplifiers. This neuromorphic architecture is a step towards exploring analog computers whose distributed signal restoration permits them to perform reliably sequential computations of great depth