CompreHending And Mitigating Error in AnaLog ImplEmentations of On-Die Neural Networks – CHAMELEON

Artificial intelligence, in general, and neural networks, in particular, are used in myriads of contemporary applications. While software and/or digital hardware implementations of neural networks currently enjoy the lion's share of the market, a number of emerging realities are necessitating the development and deployment of analog on-die neural networks. Specifically, the exponential growth of sensory data from world-machine interfaces, known as the analog data deluge, along with the power, area and response-time constraints of distributed edge computing systems, call for the ability to autonomously sense, perceive, reason and rapidly act. This ability, which avoids overwhelming communication, storage and computational infrastructure, is promised by analog on-die neural network implementations. Such implementations, however, are susceptible to parametric differences introduced by manufacturing process variation, operational conditions variation, as well as silicon aging, thereby introducing a new set of robustness and resilience challenges for the implemented models. Accordingly, this project seeks to comprehend the impact of parametric differences on the learning capacity of on-die analog neural networks, develop error mitigation solutions and assess their effectiveness through implementation of pertinent experimentation platforms.