Scalable Analog Neural-networks (ScAN)

Background
The Defense Advanced Research Projects Agency (DARPA) is soliciting innovative proposals in the technical area of scalable, robust, and power-efficient analog neural network (NN) architectures and circuits that could directly interface with the analog outputs of conventional sensors. The proposed research should investigate innovative approaches that enable revolutionary advances in science, devices, or systems. DARPA is open to teaming when appropriate and strategic to achieving program outcomes.

Work Details
The ScAN program is a 54-month, 2-phase program with a 27-month Phase 1a (base), 12-month Phase 1b (option), and 27-month Phase 2 (option). The program will address two key Technical Challenges: TC1 - Efficiently overcoming scaling limitations and short-term performance fluctuations, and TC2 - Efficiently overcoming device dependent variations and long-term performance changes.

The program will develop techniques to address TC1 and TC2 for realization of robust, accurate, and power efficient analog NN circuits at intermediate scales in Phase 1a. In Phase 1b, the program will validate the techniques developed in Phase 1a and demonstrate processing techniques, including training of the intermediate-sized networks. In Phase 2, the program will develop and demonstrate techniques to scale analog NNs to large systems while maintaining power efficiency, accuracy, and robustness.

Period of Performance
The ScAN program is a 54-month, 2-phase program with a period of performance estimated to start in December 2024.