Can you teach a computer to dream?
As processing power leaps ahead, artificial intelligence, or machine learning, has a strange side-effect.
Creating artificial neural networks that were like a simplified version of the brain meant some of the mysteries of the human brain could appear in artificial versions of it.
University of Otago department of computer science lecturer Lech Szymanski said although machine learning could be very successful, computer scientists did not completely understand how it worked.
"We do not explicitly design a logic that does something, we create a model that is capable in theory of learning anything you want it to learn."
An artificial neural network mimicked the way connections form in the human brain.
"It is like a tiny little brain, so these deep-learning models, they are inspired by the brain, and at the same time they are really simplified versions of the brain."
One way humans learn is to try different options. For example, when playing a computer game, a human would try different routes to see if they were more successful.
Haitao Xuone, one of Dr Szymanski’s students, was working on what he called learning by surprise, Dr Szymanski said.
"The computer agent internally has this logic where it keeps track of what it experiences, and learns to understand the environment and rewards itself for seeking new experiences."
In this way it was less likely to wander aimlessly around situations it had already seen.
It could be taught one task and it could learn it very well, for example, playing one computer game, but when you then moved on to another task such as playing a different game it would forget the first task.
This was called catastrophic forgetting. Another of Dr Szymanski’s students, Craig Atkinson, was exploring a possible solution to this by utilising research first advanced in the 1990s called pseudo-rehearsal.
An addition to the system could act as a memory network, which could learn to memorise situations to be called upon at a later date.
"I could compare it to like our sleep," Dr Szymanski said.
"The network sort of goes into this other mode where it is randomly poked and it creates these ghostly images of what it remembers."
This was almost how humans relived experiences or tasks over and over in dreams.
"Dreams are quite often things that have happened to you, but also random things can pop up."
This allowed the agent to play multiple games and not forget how to play the first game, Dr Szymanski said.
"It is very analogous to what we have in the brain."
