Navigation auf uzh.ch
Navigation auf uzh.ch
We are combining neuroscience, mathematics, computer science, and engineering to develop both a theoretical formalism and its supporting technology for building spike-based neuromorphic processors.
The theoretical efforts aim to understand the role of spike-based learning in physical systems (be it natural or artificial), affected by constraints that are often not taken into account in computer simulations (such as the effect of bounded weights with limited and small resolution, or the effects of variability and drift in the network parameters). Our technology developments are based on the design of mixed-signal analog/digital electronic circuits that implement massively parallel recurrent spiking neural networks. We design innovative ultra low-power analog circuits, typically operated in the sub-threshold domain, to directly reproduce and emulate the dynamics of real neural processes. We use state-of-the-art asynchronous data-driven (self-clocked) digital circuits to manage the communication of spikes (address-events) among multiple neural populations, within single chips, and across multiple chips.
The neuromorphic processors we develop are fundamentally different from conventional processors used in standard PCs based on von Neumann architectures. They are composed of massively parallel arrays of simple but inhomogeneous processing elements in which memory and computation are co-localized. This design does not allow for the virtualization of time and the rapid transfer of partial results back and forth between the central processing units and memory banks situated outside the architecture core. Instead, our synapse and neuron circuits process input spikes on demand, as they arrive, and produce their output responses in real time. As these elements must operate with time constants that are well matched to those of the natural world signals they are designed to process, we slow down our neural electronic circuits to have time constants of the order of milliseconds.
Our theoretical and technology development efforts are aimed at the synthesis and construction of neuromorphic cognitive systems that can carry out sensory-motor tasks efficiently, interact with the environment in real time and produce intelligent behavior, rather than at the attempt to build large-scale general purpose neuromorphic computing systems as generic simulation engines or data processing architectures.
