Giacomo Indiveri , Fabio Stefanini *
A brain-inspired neuromorphic architecture for robust neural computation
Progress in Neuromorphic Engineering has led to the development of electronic circuits and systems that can efficiently emulate the properties of key biophysical elements of the nervous systems. In recent works, hybrid analog/digital electronic circuits of neurons and synapses have been used as the basic components in a range of compact and low-power VLSI devices designed to implement bio-inspired computation. It has been recently demonstrated that it is possible to implement spike-based neural processing systems that can process sensory signals in real-time and take decisions based on external inputs, internal states, and on stored memory of the recent past. However the remarkable ability of biological systems to adapt to new, unknown environmental conditions, learn from their experience, and extract relevant context-dependent information from the world, is yet unparalleled. A large spectrum of experimental evidence in biology has shown that abilities of this type emerge from a rearrangement of the synaptic connectivity in the neural network. Many theoretical studies have tried to reproduce these experimental observations, mainly via software simulations of spike-based learning in neural computational
models. We developed a set of low-power circuits for neurons and synapses emulating spike-based synaptic plasticity in hardware and used it in a task of hand-written digits recognition to benchmark the memory formation in the artificial hardware neural system. The neuromorphic system comprises hardware Integrate & Fire (I&F) neurons and plastic synapses configured as binary classifiers (or "perceptrons"). We present experimental results that characterize their classification properties, and demonstrate, with the aid of software simulations, how the proposed architecture can achieve good classification performances despite the non-negligible diversity of its
components. This approach is inspired by machine learning techniques where simple, weak classifiers are typically combined to obtain a good global classifier. We demonstrate therefore how the compact and low-power (but inhomogeneous and noisy) electronic circuits developed are in good accordance with these machine-learning theories, which require variability in the classifier responses to optimize the overall system performance.