A Brain-Inspired Approach to Rapid and Energy Efficient Information Processing: AI on the Fly
The electronics revolution of the past five decades has been pivotal to improving the quality of life of people across the globe. However, this revolution has been powered primarily by exponential improvements in silicon integrated circuits with time (known as Moore’s scaling), which has now run up against some seemingly intractable roadblocks. Without exponential increases in computing capabilities, transformative visions such as the Internet of Things, autonomous transportation, disaster resilient infrastructure, and personalized medicine will be throttled by the inability of current computing technologies to handle the magnitude and complexity of human and machine generated data. Processing, storing, and transmitting information already accounts for ca. 10% of global energy use; estimates suggest that by 2040, the demand for computation will be 10× higher than the projected global energy supply. The human brain serves as a powerful exemplar of processing complex information with minimal energy dissipation. The energy efficiency of the brain stems from its ability to use an adaptive approach, responsive to “events” triggered by external stimuli, wherein multiple internal states are accessed and retained in a manner reflective of past conditioning (the ability to evolve and “learn”). Unlike binary von Neumann architectures, where computing operations constantly fetch data from a central processor, neural elements use distributed, dynamically evolving, and interacting weights (memories) stored across a dense, highly interconnected network. Realizing solid-state analogs of neural circuitry, in what are known as ‘neuromorphic’ materials, holds promise for enabling a new energy-efficient computing paradigm. Our interdisciplinary team seeks to establish a new paradigm for harvesting, storing, analyzing, and transmitting the most valuable information by designing physical analogs of the human brain.
To emulate neurons and synapses, the electrical conductance of materials must be switched across orders of magnitude in an energy-efficient manner, thereby defining physical analogues of action potentials, i.e., neural spikes. Additionally, these spikes must be precisely correlated across space and time (to enable learning) with tunable retention of internal states (constituting memory). Furthermore, individual devices must be assembled within an interacting network that collectively provides emergent learning, memory, and processing functions. The ambitious vision of our X-grants team is to design new dynamical materials and systems that directly emulate the functionality of neurons and synapses to combine memory, computation, and communication within a single fabric. We propose to develop novel dynamical materials with nonlinear conductance switching emulative of neural elements and to build altogether new circuit elements from such materials that can perform tasks that would require hundreds to thousands of transistors. Such an advance will dramatically decrease the cost of information processing while increasing the energy efficiency by orders of magnitude. A major advantage afforded by such high-fidelity neural emulation will be the immediate updating of the system (learning) in response to external events, thereby realizing the full promise of artificial intelligence to learn and respond in real-time. The X-grant will enable the interdisciplinary team to establish an interdisciplinary track record (bridging materials to circuits to systems) of high-profile publications and to build a programmatic infrastructure that showcases the credentials of Texas A&M University to sustain an activity of this vast scope and importance. Our X-grant activities will furthermore spur curricular transformation, innovation, and development of a campus research hub that will connect and elevate research activities across different colleges to enable unprecedented utilization of artificial intelligence spanning a range of problems.