IBM Presents A Method To Make AI Learn From Its Memory

Computers are light-years ahead of humans in term of computational prowess and logic skills. When we program computers into using its amazing ability into learning, needless to say, once it has learned something it’d turn the best of us into sore losers. We have had Artificial Intelligence (AI) programmes so good at playing video games like StarCraft they effortlessly beat pro players.

But they can only do it if everything’s done by the rules. When the rules change and the AI is left in an unfamiliar situation, it is not as good adapting like us. AI trained to play Pong can’t even handle a tiny change like a shift in distance between the paddles.

But that’s about to change.

The Adaptable AI

An IBM research presenting at an AI conference in May promised a way we could give AI adaptability. They will give the computer an algorithm that can improve itself as it functions, using its virtual ‘memory’ to adapt to new parameters and environment without necessitating re-training.

If we were to use the Pong AI example, this means the AI can adapt to the minuscule shift in paddles’ distance and play the game with the same perfection as it did when there was no change. This is according to the research paper published by IBM-Watson AI Lab.

It is surely a step forward in term of AI’s flexible reasoning, and it is only a prelude to what the future could hold for this exciting tech.

'Catastrophic Forgetting'

Though it may sound scary and foreboding to some, the ultimate goal of projects such as this one is to build artificial general intelligence, or ‘strong AI’. The type you’re used to seeing in science fiction that talks, expresses, and acts like a human being. For sure, we’re still a long time from seeing this dream through into reality, but the research put forward an interesting idea of where to begin. It allows AI to learn and adapt flexibly, mimicking the way human’s brain updates, processes information in real time.

Matt Riemer - an IBM scientist told the tale of the ‘forgetful AI’ in one of his unofficial blog posts that had yet gone through any sort of peer review. He described the problem when an algorithm is struck by what he called “catastrophic forgetting”, where the AI forgot all of its training and returned to being a blank slate as soon as they’re trained in doing a new task.

Other teams of researcher have tried to fix this problem. At Google’s DeepMind, they even tried to give the AI something that’s similar to the imagination that could allow it to store ‘memories’ better. What IBM is trying to do here is just like that, except it’s done from a different angle. The final goal is to create an AI that can learn on its own and adapt independently without the needs of human supervision.