BMBF Bernstein Group "Components of cognition: from small networks to flexible rules" Individual configurability of plastic synapses in neuromorphic VLSI

The pioneering work of C. Mead \cite{Mead89} has introduced the term ``neuromorphic engineering'' for a growing family of analog, sub-threshold circuits, which implement the accepted equivalent circuits of biological neurons and synapses in VLSI technology. The ultimate aim of neuromorphic engineering is to mimic the capabilities of biological perception and information processing with a compact and energy-efficient platform. We believe that this goal necessitates from the outset some mechanism of ``learning" that enables neuromorphic devices to adapt (or re-configure) themselves while interacting with an environment. Emulating the example of biological neurons and synapses, our neuromorphic devices attain an ability for "learning" by incorporating ``Hebbian-like" mechanisms of synaptic plasticity. In the "Hebbian" scenario, the efficacy of a synapse is enhanced (i.e., its impact on the post-synaptic neuron is increased), when the activities of pre- and post-synaptic neurons are correlated on a suitable time-scale, and reduced if the activities are anti-correlated on this time-scale. Whether ``Hebbian" learning is based on average firing rates or on individual spikes (``spike-time-dependent plasticity", or STDP) is a matter of continuing debate and a choice that strongly influences alternative designs of neuromorphic synapse circuits. The synaptic circuits described here represent a compromise, in that they combine rate-based ``Hebbian" learning with many aspects of STDP. We illustrate tests and measurements performed on an analog, VLSI chip implementing 128 integrate-and-fire (IF) neurons and 16,384 plastic synapses. Each synapse may be individually configured to be either excitatory or inhibitory and to receive either recurrent input from an on-chip neuron or AER-based input from an off-chip neuron.