Saturday, February 9, 2008

On Intelligence

One of the greatest obstacles in the path of re-creating human intelligence is defining just what is it that we are trying to reproduce. Defining intelligence remains an elusive goal, but I think Jeff Hawking's book On Intelligence marks a step in the right direction. It makes the point that, since the human brain (or rather the neocortex) is the only truly intelligent device we know, to find out what intelligence is we must discover just what it is that brains do that make us intelligent. Supported by the great body of knowledge on neurology accumulated over the years, he goes on to describe his "memory-prediction framework" (MPF) theory, which states that intelligence is the ability to derive future (or unobserved) events from past experiences.

Now, if you reacted anywhere close to how I did when I was first introduced to it, you probably thought something along the lines of "that's ridiculous", but please bear with me. In our brains, the neocortex is the part responsible for the kind of adaptive skills we usually relate to intelligence. A multi-layered sheet of neural tissue covering the primitive ("reptile") brain, it is only present in mammals, and is most developed in humans (other mammals have bigger brains, but their neocortex are smaller relative to body size, or contain less than our 6 layers). Hawkings argues that the neocortex's basic function is creating "invariant representations" from sensory input streams: when combined with live streams, these invariants can recreate past experience – but they can also be matched against new or incomplete sensory data, guiding the brain into "filling in the blanks" in the new or incomplete stream.

How does this relates to intelligence? Imagine for a moment you are a little child, and you love cookies. One day you observe that your mother takes a huge jar from the top of the fridge, reaches into it and gives you a cookie. Then you imagine that, if you climbed up the fridge and took that jar, you too would be able to reach into it and get yourself a cookie. From observation, your brain derived an invariant representation – the sequence of steps needed to get a cookie; next, you filled the invariant with a parameter – yourself – and thus predicted the natural consequence of your acts – by climbing the fridge, you could help yourself to another cookie.

One of the features of this definition is that it is completely detached from behaviour: you'd still be an intelligent child (perhaps even more so :P) if you just stood there, without acting on your predictions. It can also easily encompass other definitions, for example:
  • "Intelligence is the ability of solving problems": by observing enough examples of "problems" and their "solutions", an MPF-enabled agent can work an invariant representation of the relation between them, so that, when it is later introduced to a new "problem", it can derive the related "solution";

  • "Intelligence is the capability of a system to adapt to its environment and to work with insufficient knowledge and resources": MPF's invariants allow for easy generalization over vast problem classes; they also allow agents to derive ("predict") information about the environment from incomplete, or even partially incorrect, input sets, and then use the treated input to work an appropriate answer.
So is MPF the One True Definition of Intelligence™? Hawkings himself cautions that some parts of his theory are bound to change as it is more deeply studied, even though he is confident the basic idea is right. If you ask me, "all models are wrong, some models are useful" – and at least for me, the MPF points to some very promising directions.