Sunday, 15 February 2015

Evolution and inductive inference

Inductive inference refers to using knowledge to make predictions. It's the basis of the scientific method. This was once famously disputed by Karl Popper - but is now widely accepted.

Sequence prediction is a well-known type of induction problem. For example, what comes next: 3,4,7,11,? Inductive inference relies on knowledge acquisition - i.e. learning. Inductive inference is inherently fallible and probabilistic.

It is clear that evolution builds systems capable of performing inductive inference - namely animal brains. However, there's another link between evolution and induction - since both processes involve using knowledge of the past to make predictions about the future. This is perhaps not obvious, but if you think about it, every organism represents a kind of prediction about the environmental conditions it will encounter. The prediction is that the organism will encounter an environment have allows it to reproduce.

Many have linked evolution and inductive inference in this way. For example, in

Universal Darwinism
, John Campbell wrote:

The Darwinian process may be the only physical mechanism known to science capable of accumulating knowledge from experience. It performs inference and is a physical analogue of Bayesian updating.
Similarly, in Probably Approximately Correct, Leslie Valiant wrote:

To see evolution as a form of learning, we view the genome in evolution as corresponding to the hypothesis in learning.

Philosophers of science who view scientific knowledge acquisition as a form of Darwinian cultural evolution are also implicitly making the same link between Darwinian evolution and inductive inference.

So: what is the link between evolution and inductive inference? I have a couple of comments to make:

  1. The claim that Darwinian evolution is a type of learning (made by Leslie Valiant in Probably Approximately Correct), isn't really right. Darwinian evolution can also produce genetic drift - which has little to do with learning. Evolution sometimes results in knowledge acquisition and successful inferences. Other times it results in progressive knowledge loss and extinction. It depends.

  2. It is tempting to link fitness in evolution with scientific truth or accurate knowledge. However, this association is inferior in practically every way to linking fitness with popularity. Phlogiston and the aether are popular mistaken ideas. They are kept around to help show where not to tread. Death in evolution maps poorly on to falsification in science.

The idea that Darwinian evolution underlies most systems that perform inductive inference is important and under-appreciated. Those engaged in creating machines that perform inductive inference tend to associate Darwinism with genetic algorithms. Those are often seen as being just one tool in a large toolkit. They are generally used in those cases where the only thing you know about your solution space is a scalar quality metric. Knowledge of memetic algorithms paints a rather different picture. In fact, evolutionary algorithms are fundamental.

No comments:

Post a Comment