Inspired by the way sleep helps us fix the knowledge accumulated during the day, researchers have enabled an artificial intelligence to learn how to perform new tasks.
Training periods interspersed with sleep phases
Most of AI usually master only a limited set of tasks, and cannot acquire new knowledge without forgetting everything they have learned previously. In order to overcome this limit, an international team of scientists taught an artificial neural network to master two different tasks without “overwriting” previously established connections, via targeted training and sleep periods (simulated by activating neurons in a particular pattern).
First, the team tried to form the neural network to the first task, then to the second, before adding a period of sleep at the end, but quickly realized that this sequence systematically erased the connections related to the initial task.
” Complementary experiments showed that it was important to quickly alternate training and sleep sessions while the AI learned the second task “, explains Erik Delanois, researcher at the University of California and lead author of the study, published in the journal Computational Biology PLOS. ” This helped consolidate initially memorized connections, which would otherwise have been erased. »
This method notably taught two different foraging patterns to the neural network, which was able to identify simulated food particles while avoiding their toxic counterparts.
Towards more versatile impulse neural networks
The team explained that they used “pulse” neural networks, which consume much less energy than conventional devices, but whose complex design, inspired by living things, currently limits their use to very specific domains and tasks.
” The goal of AI’s lifelong learning is to have the ability to combine different experiences in an intelligent way and apply that learning to new situations, as animals and humans do. “Says Hava Siegelmann, from the University of Massachusetts.
” The current trend is to bring ideas from neuroscience and biology to improve existing machine learning, and sleep is one of them. concludes Maxim Bazhenov, from the University of California.