Can machines teach themselves about time?
To expand the shelf life of artificial intelligence systems, the Defense Advanced Research Projects Agency is looking for researchers to help it build a neural network architecture with a self-understanding of time.
The Time-Aware Machine Intelligence (TAMI) program aims to develop new architectures that introduce a meta-learning capability into machine learning that would allow a neural network to learn to teach itself about the time dependencies within its encoded knowledge.
Currently, neural networks do not take the implications of passing time into their encoded knowledge, so outdated or finite information may not necessarily become less relevant, and decisions based on knowledge lacking a time dimension may be erroneous. Additionally, correcting this problem requires frequent and costly retraining to optimize performance.
“TAMI’s vision is for an AI system to develop a detailed self-understanding of the time dimensions of its learned knowledge and eventually be able to ‘think in and about time’ when exercising its learned task knowledge in task performance,” DARPA wrote in a presolicitation.
Rather than modeling time properties into the source data and encoding it in a neural network, TAMI “focuses on modeling the time property of its own learning.” In other words, DARPA said, “TAMI will develop a form of meta-learning into neural networks.”
Such capability would give AI inference systems a longer life without performance degradation, even when some of their encoded knowledge becomes obsolete. It would enable time-based causal awareness for machine learning models and the choreography of actions between different machine learning agents. Ideally, it would also support the “mental time travel” AI would need to recall past experiences to inform present or future decisions.
A live question and answer session with the TAMI program manager will be held on polyplexus.com on Sept. 29.
Read the full solicitation here.
Connect with the GCN staff on Twitter @GCNtech.