Abstract: Neural networks, learning algorithms and their hardware accelerators have emerged as a successful technology for solving a wide range of complex pattern recognition tasks. However training and running such algorithms often requires the use of very large data-sets, heavy computational loads, and very large power budgets.
Standard computing technologies will not be able to sustain the increasing demand for intelligent processing of data and signals measured "at the edge", i.e., in the environment or in areas that are not connected to data centers for central processing. One promising alternative to solving this problem is represented by the "neuromorphic engineering" approach. Following this approach we have developed mixed-signal analog/digital neural processing circuits that enable the construction of ultra low-power brain-like computing architectures. I will present examples of such circuits and architectures, showing how they can be used to carry out signal processing tasks in real-world application scenarios. I will also show how the constraints imposed by the use of these brain-like computing elements led to the definition of computational primitives and architectural solutions that on one hand can contribute to understanding how real neural processing systems work, and on the other can lead to alternative forms of computation dramatically different from the Turing-machine paradigm.
Bio: Giacomo Indiveri is an engineer by training, but he has always been interested also in physics, computer science, and neuroscience. He combines these disciplines by studying real and artificial neural processing systems, and by building hardware neuromorphic cognitive systems, using full custom analog and digital VLSI technology.