Sleep helps consolidate memories and this also applies to artificial neural networks by allowing them to learn multiple tasks at a time
When our body rests , the brain has the time it needs to reinforce learning and deposit the memories acquired during the day. This is also true for artificial intelligence according to a study published in PLoS Computational Biology. The artificial equivalent of sleep prevents artificial intelligences from losing newly learned knowledge.
Many of these artificial intelligences can become hyperspecialized in a single task, repeating the same training procedure over and over again on the available data, but they can only master a specific set of information at a time. This means they can’t add more knowledge without losing what they learned before, which is a problem for mimicking the human continuous learning process , where memories become layered and you only need to look at something once to have a memory of it too. imprecise in mind.
Pavel Sanda and colleagues at the Institute of Computer Science of the Czech Academy of Sciences in the Czech Republic have instructed a particular type of artificial neural network to learn two different tasks without erasing the first learned task. This was possible by including periods of simulated sleep in the training , stimulating the artificial neurons in a chaotic way, as happens in the brain while you sleep.
However, the scientists ensured that in its disorder, the activity retraced the neural activation sequence of the learning section , causing that kind of replay which, in sleep, allows us to consolidate memory in our brain. This process helped the AI consolidate connections learned in the first task and accumulate new ones, but only when sandwiched between quick learning sessions. If the rest phase is inserted at the end, i.e. after the two learning phases , what was learned in the first task has been lost.
- Sleep is also good for artificial intelligence (focus.it)
