Memristor-based neural network circuits
Abstract
Artificial intelligence-at-the-edge (Edge AI) is the key to the extension of AI services into the domain that requires stringent privacy/security, e.g., medical and military applications. Realization of the Edge AI is in need of dedicated hardware that remarkably accelerates neural network calculation at a smaller cost of power and energy than general-purpose hardware such as central processing units and graphics processing units. The research and development trend in dedicated hardware to the Edge AI explains digital hardware based on conventional complementary metal-oxide-semiconductor (CMOS) technology as a short-term solution. Yet, long-term solutions should outweigh the CMOS-based hardware with regard to important figures of merit such as power and energy consumption, neural network calculation speed, areal density of neural networks, and so forth. To this end, we propose memristor-based spiking neural network (SNN) circuits embedding on-chip learning capability. Hereafter, we refer dedicated hardware to SNN implementation as neuromorphic hardware and computing using such hardware as neuromorphic computing. Memristors exhibit diverse types of memory effects with wide-range of memory retention time, ranging from a few tens of milliseconds to a few years. Such rich dynamics are the heart of neuromorphic hardware, which are conventionally realized at the circuit level at the cost of large number of transistors and/or large capacitors. Given that memristors are simple metal/insulator/metal structure and freer from scaling issues than conventional circuit elements, memristor-based neuromorphic hardware may be the ultimate solution to dedicated hardware to the Edge AI at the maximum density. We plan to approach memristor-based neuromorphic hardware at different levels: (i) materials, (ii) circuit, (iii) network integration, and (iv) learning algorithm level. The proposed multi-level research will develop complementary technologies over the different levels, which accelerates technology development with the least trial and error. Particularly, memristor-based neuromorphic hardware as dedicated hardware to the Edge AI comes into its own with embedded learning algorithm that is essential to the true Edge AI independent of the Cloud.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- May 08, 2020
- Source ID
- N629092012021
Entities
People
- Doo Seok Jeong
Organizations
- Office of Naval Research
- United States Navy