DARPA wants AI that can learn from the experiences of others


Humans progress faster when we learn from the experience of others, and scientists at the Defense Advanced Research Programs Agency, or DARPA, want to translate that into lifelong learning models for artificial intelligence.

The research agency has opened a new artificial intelligence exploration opportunity to fund work on “the technical field of lifelong learning by agents” – AI systems – “who share their experience with each other, ”according to an ad on SAM.gov. DARPA is offering up to $ 1 million per proposal through the Shared Experience Lifelong Learning Program, or ShELL.

“Lifelong learning is a relatively new area of ​​machine learning research, where agents continually learn as they encounter varying conditions and tasks when deployed on the job. field, gaining experience and knowledge and improving performance on new and previous tasks, ”says the funding announcement. .

The review details how this differs from traditional ‘train then deploy’ machine learning, which tends to fail in three ways:

  • Unpredictable results when entry conditions not representative of training experiences are encountered.
  • Catastrophic forgetting of previously learned knowledge useful for new instances of previously learned tasks.
  • The inability to perform new tasks effectively, if at all.

While lifelong learning is not a new concept for AI research, the announcement notes that current research has focused on learning models for individual systems, rather than on “Populations of LL agents who benefit from each other’s experiences.

Under the ShELL program, DARPA will fund projects that start with a large number of identical AI systems that are then deployed in different real-world situations. As individual systems adapt to their environments and tasks, the information gathered will be shared with the entire group, improving training data for all.

This differs from other mass AI training programs where a group of systems work together to accomplish a single task and learn a shared set of lessons.

“Shell is not a distributed learning framework that assumes a breakdown of tasks and training data / experiences purely for training effectiveness or due to external policies restricting the combination of source data sets,” states the financing notice. “In contrast, ShELL rewards individual agents based on their performance in their own tasks using lessons learned from their own actions combined with those learned from other agents.”

DARPA officials identified three major challenges that proponents must address in their offers:

  • Content: What knowledge should be shared and incorporated and which should be ignored?
  • Communications: when and how should knowledge sharing take place?
  • Compute: Ensure learning groups have sufficient compute power through a combination of edge and cloud resources.

The project will be carried out in two phases, with grants of up to $ 1 million per proposal. Phase I focuses on a six-month feasibility study, with financial support of up to $ 300,000. Projects moving to Phase II will develop a 12-month proof of concept, with maximum funding of $ 700,000.

Allocations will be made using the other DARPA transaction authority.

The ShELL program is now accepting submissions, with the goal of awarding prizes by September 24.


About Author

Leave A Reply