Why AI needs to get a good night's sleep as well

Humans function better after having a good night's sleep. It seems that it's the same for AI as well!

A lesser known fact about some types of AI models is that they can suffer from a phenomenon called Catastrophic Interference, sometimes also known as Catastrophic Forgetting. As the name implies, this is when AI can forget already learned information, and it's a big problem to overcome.

The phenomenon happens because of sensitivity to new information in distributed neural networks. When new information is introduced, it can affect the weights that have already stored learned information, and so 'prior knowledge' can be affected destructively. In short, when new patterns of information are received, the AI loses the ability to recognise the old patterns as well.

The problem was first recognised in the late 80's by Michael McCloskey and Neal Cohen while performing two different experiments with backpropagation neural network modelling. Since the issue was noticed, several methods of solving the problem have been put forward, but the latest idea is one that at first seems a little weird. Researchers at the University of Catania, Italy, have developed a way for AI to mimic sleep.

Now, the use of the term 'sleep' is more of a concept than tucking up the computer in bed and giving it a mug of Horlicks. Humans use sleep for a multitude of reasons, one of which is the consolidation of memory and learned information. The researchers at the university have developed a new learning model called wake-sleep consolidated learning (WSCL), and it attempts to mimic the way humans reinforce information during sleep cycles.

With WSCL learning works by training the AI on new data during its 'awake' cycle, but then it will enter a 'sleep' phase whereby it will go through the data it learned during its awake phase and a highlight reel of previous lessons. The idea is that the AI gets to reinforce data into a kind of long term memory by making sure it gets reminded of old data, preventing any risk of it being overwritten by new information.

Dream a little dream of me

However, the sleep behaviour mimicking doesn't stop there. The WSCL learning method also encompasses a 'dream' phase, too. This works by merging together previous concepts and datasets. For example, if the AI has learnt how to recognise different animals, the dream phase feeds the AI combination data, for example a bird mixed with a fish. The idea is that it merges previous paths of 'neurons', and will assist the AI in learning new concepts at a later time, forcing the AI to learn ever more complex patterns.

The big question is, of course, does the method work? Well, according to Concetto Spampinato, who headed up the university research, AI models trained using the method were between 2-12% more accurate at identifying at identifying the contents of images, as well as 'remembering' how to perform older tasks more reliably than a traditionally trained system. The training method can be applied to all existing AI systems, too.

The development could be very significant, not just for the usefulness of AI systems like ChatGPT, but also for any potential AI driven vehicle, which might need to learn and adapt to different scenarios and situations. Whilst ever increasing processing speed and complexity will make AI grow exponentially in capability, the fact that it is possible to make all existing AI systems use the method means that in combination with other progress, AI might begin advancing at an even greater pace than we're used to.