DARPA – Proy ShELL, busca que las computadoras compartan experiencias entre sí, el aprendizaje automático es un área relativamente nueva de la investigación informática, en la que las computadoras aprenden continuamente a medida que se encuentran con diferentes condiciones y tareas mientras se implementan en el campo.
ARLINGTON, Va. – U.S. military researchers are asking industry to devise a new kind of artificial intelligence (AI) computer programming that enables computers not only to learn from their experiences, but also to share their experiences with other computers.
Officials of the U.S. Defense Advanced research Projects Agency (DARPA) in Arlington, Va., issued an artificial intelligence exploration opportunity on Monday (DARPA-PA-20-02-11) for the Shared-Experience Lifelong Learning (ShELL) project.
ShELL seeks to advanced computer sciences in lifelong learning by computers that share experiences with each other. Lifelong learning is a relatively new area of machine learning research, in which computers continually learn as they encounter varying conditions and tasks while deployed in the field.
This differs from the train-then-deploy process for typical machine learning systems, which often results in unpredictable outcomes; catastrophic forgetting of previously learned knowledge; and the inability to execute new tasks effectively, if at all.
Current lifelong learning research assumes one independent computer that learns from its own actions and surroundings; it has not considered populations of lifelong learning computers that benefit from each other’s experiences.
The total award value for the combined phase-one base and phase-two option is limited to $1 million per proposal.
Algorithms used for lifelong learning typically require large amounts of computing resources, including server farms, graphics processing units (GPUs), and other resource-consuming hardware, and typically do not have to address communication resource limitations.
The Shared-Experience Lifelong Learning (ShELL) program extends current lifelong learning approaches to large numbers of originally identical computers. When these computers are deployed, they may encounter different input and environmental conditions, execute variants of a task, and therefore learn different lessons.
Other computers could benefit if one computer could share what it has learned with the other computers. Such sharing of experiences could reduce the amount of training required by any individual computer.
ShELL is distinct from approaches that reward a federation of computers for collaborating or competing on a common global task, either by dividing the task into pieces, by assembling alternative approaches to the same task, or by evolving specialized roles.
ShELL rewards computers individually according to their performance on their own tasks using lessons learned from their own actions combined with those acquired from other computers.
ShELL has three core challenges: what knowledge should be shared and incorporated; when and how should computers share their knowledge; and develop lifelong learning algorithms that account for the size, weight, computing, and communications constraints of the platforms supporting each learning computer.
