Intuition and neural networks

Human intuition behaves similar to a neural network software algorithm: it produces an estimation of a certain probability, based on prior experience.

The same with an artificial neural network, human intuition needs to have sufficient training experiences on that field so it can extract regularities. The training experiences should be able to provide adequate and rapid feedback that can be used to adjust the internal model, for both biological and software neurons.

The domains where intuition and neural networks work well are the ones that are regular enough, so that the modeled behavior can be inferred from prior experience. Domains where there are rare discontinuities cannot be well modeled if the training experience does not contains enough of them (think global economic crises in a life time). Same, when the outcome of an action can be observed after a long time, the training feedback is inefficient, because in the meantime there may be a lot of other influences that contributed to the final outcome.

It is very hard to prove that a certain economic measure had a good outcome, because each subsequent effect is the result of many other changes in the environment. It is easy to implement an economic measure and believe it helped, when actually the outcome would have happen anyway from other reasons. Also, a good economic measure that actually helped to slow down an accelerating crisis may seem like causing the remaining effects of the crisis.

Bottom line: don't always trust your intuition, unless it was verified in many similar situations. Use your reason also!

No comments:

Post a Comment