Neural networks and deep learning algorithms are currently achieving impressive state-of-the-art results on computing tasks that operate on stored data sets. However, artificial computing systems are still vastly outperformed by biological neural processing ones for tasks that involve processing of sensory data acquired in real-time in complex and uncertain settings, and closed-loop interactions with the environment. This difference is remarkable especially when size and energy consumption are factored in. One of the reasons for this gap is that, as opposed to conventional computing architectures, in biological neural systems  computation is tightly linked to the to the physics of their computing elements and to their temporal dynamics. In this talk I will present hybrid analog/digital microelectronic circuits that use their physics to directly emulate the biophysics of the neural processes and memory elements they model. I will demonstrate examples of brain-inspired architectures that integrate massively parallel arrays of such circuits to implement on-chip on-line spike-based learning and computation, and will describe the advantages and disadvantages of these types of computing architectures compared to conventional computing systems. I will argue that the circuits proposed represent a promising approach for building intelligent and energy-efficient autonomous cognitive agents that need to process input data as it arrives, in real-time, without having to use eternal memory
storage.