The goal of this project is to develop AI algorithms and architectures that can adapt and learn after deployment. In the vast majority of AI and machine learning approaches, learning takes place only when the system is undergoing training, so it is ​“frozen” in place once deployed. Having systems that can continuously learn is critical for a wide range of applications, from autonomous vehicles to control and robotics, smart sensors, or expert systems for science applications.

In this project, we borrow from neuroscience — in particular, the insect brain — to explore how to extend learning into the deployment phase, demonstrating systems capable of lifelong learning. Such systems are characterized by the following (1) continual learning — including the ability to learn an evolving stream of tasks without distinct training and testing phases and to adapt to changes in their environment; (2) transfer and adaptation — meaning system performance in novel and known tasks improves with experience; and (3) sustainability — learning continues for an extended period using limited resources.

This project will not only develop more adaptive AI, but it will also help researchers understand how to develop new hardware that is capable of carrying out lifelong learning at the edge and help evolve new generations of brain-inspired computing hardware. For instance, we have explored the implementation of some of our algorithms in FPGAs, neuromorphic chips such as Intel’s Loihi, and novel architectures based on emergent materials.