Neuro‐Inspired Distributed Deep Learning

Professor Pulkit Agrawal, an affiliate of the MIT Computer Science and Artificial Intelligence Lab (CSAIL) and the MIT Laboratory for Information and Decision Systems (LIDS), leads this Multidisciplinary University Research Initiative (MURI) project. Agrawal's team, which includes Ila Fiete and Aude Oliva of MIT as well as researchers from Harvard University and the University of California at Berkeley, proposes an alternative to the mainstream machine-learning practice of condensing large datasets into the weights of deep neural network and discarding the training data itself. Such an approach has fundamental limitations when it comes to lifelong learning and the associated questions of generalization, long-term reasoning, and catastrophic forgetting. As such, the proposal suggests avoiding compressing data ahead of time and instead combining data on-the-fly for the environment or task encountered by the agent, using memory retrieval to improve generalization.

The work aims to articulate a set of high-level computational principles for the design of memory systems, leveraging knowledge about how the brain encodes and retrieves information from memory. It aims to determine how these principles can be leveraged to tackle challenging machine learning tasks, understand how biological memory systems represent and retrieve naturalistic inputs, and help in the integration of AI into a wide variety of real-world systems. Ideally, the end result will yield practical algorithms for generalization to new tasks, lifelong learning without catastrophic forgetting, and transfer across sensory modalities.