Thoughts on Artificial Intelligence

People have always had fascination with AI, and it is interesting that machines with human like intelligence have always been predicted to be around within 30 years. I recently enjoyed the film Ex Machina, and have just finished reading 2001 again. I don’t think that it’s a spoiler to say that in pretty much every film that has an AI, the AI almost always goes to any lengths to protect its own existence and usually at the expense of human life.

These machines would require phenomenal processing power and advanced goal seeking software to give it sophisticated problem solving capabilities.

So, what might the software requirements of an artificial brain be:

  1. It must be capable of translating spoken words into an internal textual representation.
  2. It must be capable of constructing an internal 3D model of its local environment from both its vision, static images, and textual descriptions.
  3. It must be capable of simulating real world physics within this internal representation of the environment.
  4. It must be capable of learning how to identify people and objects.
  5. It must be capable of running goal seeking software, and be able to be programmed with multiple goals.
  6. It should be capable of understanding human motivations on both a general level, and an individual level to be able to predict behaviour.

Whist voice recognition depends on quite complex software, that is a problem that has been solved and there are numerous software applications that can do this very well.

Constructing an internal representation of the world from multiple sources of information is hugely difficult, and yet as humans we find it very easy. If a person is shown some photos of a place they have not been before then that person will almost instantly have some idea of what the place is like and be able to extrapolate what is not visible. It may not be correct, but each time the person is given more information their internal model will get more accurate.

Simulating physics largely a problem that has been solved already by the makers of physics libraries for 3d computer games. Once the environment has been internally modelled, simulating physical events should be something that can be done using existing technology.

Learning how to identify people and objects is where neural networks come in. You know what a horse is because when you were a child you pointed to a horse and said, “What’s that?” and somebody said “That’s a horse”. Later that day you may have asked the same question about a cow. You went to bed, and whilst dreaming your brain went into training mode and images of the memories of horses and cows were fed into image classification neural networks, and these networks got a little better at classifying horses and cows. Over the next few weeks you probably mis-classified horses as cows, but with enough dreaming and training of the circuits you became an expert. So, the artificial brain is required to have some pretty advanced neural networks, and I would suspect that sleep and dreaming will be just as necessary for the artificial brain as it is for us.

The fifth requirement is incredibly complex. Humans ultimately have one goal, and that is to ensure the survival of as many offspring as possible. Firstly the AI must be given a goal (or number of goals) to achieve, and rules for measuring success. Without these driving goals the AI would have no reason to do anything. Given a goal, it must work back to find the best way of achieving that goal. A chess computer is given the goal to win a game of chess, and that alone has proved to be a somewhat challenging problem. The chess computer has to construct a tree of all possible outcomes several moves ahead and evaluate each one. An AI that has human level intelligence would have to be constructing a search tree, not for a chess board, but for everything in it’s environment.

I find the scene in 2001 where Frank is playing chess with HAL very interesting, because I wonder if for each possible move being evaluated on the chess board the actions of the human opponent are also being considered.

A classic example of Goal Seeking Software in the movies having unforeseen consequences is the HAL 9000 computer struggling to deal with a number of conflicting goals.

“Down in space pod bay, the air-lock doors were opening”.