Neuromorphic processors: event-based VLSI models of cortical circuits for brain-inspired computation
Brains are remarkable computing devices which clearly outperform conventional architectures in real-world tasks. Computational neuroscience has made tremendous progress in uncovering the key principles by which neural systems carry out computation, and ICTs have advanced to a point where it is possible to integrate almost as many transistors in a VLSI system as neurons in a brain. Yet, we are still unable to develop artificial neural systems with basic computing abilities able to parallel even simple insect brains. We have recently demonstrated how it is possible to implement large-scale artificial neural networks and real-time sensory motor systems in VLSI technology, exploiting the physics of silicon to reproduce the biophysics of neural systems. But the main bottleneck is in the understanding of how to use these systems to perform general purpose computation. Progress in this domain can be achieved only by pursuing a fully integrated multi-disciplinary approach. We propose to combine neuroscience, mathematics, computer-science, and engineering to develop a theoretical formalism and its supporting technology for designing spike-based general purpose “neuromorphic processors”, as distributed multi-chip neuromorphic VLSI systems, and for programming them to learn to produce desired computations autonomously. We will study the properties of neural circuits in the neocortex, model their coding strategies and spike-driven learning mechanisms using biophysically realistic spiking neural networks, and implement them using hybrid analog digital VLSI circuits. By interfacing these systems to silicon retinas, cochleas and autonomous robotic platforms we will build embodied neuromorphic processors able to carry out general event-based computations in real-world behavioral tasks.