The advanced research agency is putting up $1 million per project to overcome challenges associated with AI systems learning from each other.
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 to lifelong learning models for artificial intelligence.
The research agency opened a new Artificial Intelligence Exploration Opportunity to fund work on “the technical domain of lifelong learning by agents”—AI systems—“that share their experience with each other,” according to an announcement on SAM.gov. DARPA is offering up to $1 million per proposal under the Shared-Experience Lifelong Learning, or ShELL, program.
“Lifelong learning is a relatively new area of machine learning research, in which agents continually learn as they encounter varying conditions and tasks while deployed in the field, acquiring experience and knowledge and improving performance on both novel and previous tasks,” the funding announcement states.
The notice details how this is different from traditional “train-then-deploy” machine learning, which tend to fail in three ways:
- Unpredictable outcomes when input 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 execute new tasks effectively, if at all.
While lifelong learning is not a new concept for AI research, the announcement notes current research has focused on learning patterns for individual systems, rather than “populations of LL agents that 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 the individual systems adapt to their environments and tasks, the information gathered will be shared with the entire group, improving the training data for all.
This differs from other mass-AI training regimens in which a group of systems work together to complete a single task and learn a shared set of lessons.
“ShELL is not a framework for distributed learning that assumes task and training data/experience decomposition solely for training efficiency or because of external policies restricting the combining of source datasets,” the funding notice states. “In contrast, ShELL rewards agents individually according to their performance on their own tasks using lessons from their own learned actions combined with those acquired from other agents.”
DARPA officials have identified three major challenges for proposers to address in their bids:
- Content: What knowledge should be shared and incorporated and what should be ignored?
- Communications: When and how should knowledge sharing occur?
- Computation: Ensuring learning groups have enough computing power through a mix of edge and cloud resources.
The project will be completed over two phases, with awards topping out at $1 million per proposal. Phase I focuses on a six-month feasibility study, with up to $300,000 in funding support. Projects that make it to Phase II will develop a proof-of-concept over 12 months, with maximum funding of $700,000.
The awards will be made using DARPA’s other transaction authority.
The ShELL program is accepting proposals now, with the goal of making awards by Sept. 24.