University of Colorado, Boulder - Associate Chair for Learning, ECEE Dept. / Assoc. Professor by Courtesy, Dept. of Physics / Fellow-Renewable & Sustainable Energy Institute (RASEI)
"Neuromorphic Computing Devices: Toward Hardware-Implemented Parallel Distributed Processing"
Abstract: CMOS integrated circuits have followed Moore’s Law to an astonishing degree over the last several decades, but the end is in sight as the length scales of devices enter the sub-10 nm regime. While today’s computing machines are sufficiently powerful to carry out typical tasks in information processing and communications, they are far away from enabling ever-more demanding tasks such as real-time classification and decision making in complex environments, adaptive learning, and other higher cognitive functions that we might broadly associate with artificial intelligence. High performance machines such as IBM Watson or Google’s AlphaGo are capable of complex deep-learning tasks; however, they consume several megawatts of power, are optimized for specific tasks, and are not portable. In contrast, the human brain consumes ~20 watts of power, is undoubtedly more versatile as a cognitive instrument, and of course is quite easy to move around.
The emerging field of neuromorphic computing aims to develop neuron-like processing units that are implemented directly into hardware. Such systems are capable of meaningful function with just a few simulated neurons, but they can also be massively scaled to perform complex deep learning and AI tasks with millions of neural units. Industrial development of such technologies is already rapidly underway. For instance, IBM has developed a neuromorphic chip, called True North, that simulates ~1 million neurons with 256 million synapses, yet uses only 70 mW of power. In this talk I will overview the function of neural implementations, including giving an example of their ability to overcome degradation in a biologically-realistic network. I will then discuss various materials and devices with which neuromorphic circuits can be implemented, including memristive devices based on metal oxides or conducting polymers, small neural circuits based on semiconducting polymers, and newly emerging organic electrochemical transistors (OECTs) with which several basic hallmarks of neural function have been demonstrated.
To view the full schedule, please visit: http://physics.mines.edu/PH-physics-colloquia