Thursday, April 24, 2008

It Is Good To Be Wrong

The brain by definition and experiment is a neural network.

It is programmed by interaction with the environment. You feed it patterns along with feedback from the environment. Some patterns you give positive scores. Some patterns you give negative scores.

If you are never wrong (an attitude some have but an impossibility in fact) your neural network no longer matches your actual environment and you will keep making the same mistakes over and over.

The best thing to do with mistakes is to admit them as quickly as they are recognized. It will readjust your pattern recognition system and future decisions will be better. Pain and error are essential features of life. The early admission of error and embrace of pain improves the odds of future avoidance of error and pain.

Which is why it is good to be wrong.

2 comments:

tomcpp said...

Unfortunately you're wrong. It's not good to be wrong. It's good to actually BE always right.

Why after all, would a brain need to dream ? When people tell their dreams they're always "wrong". They don't follow physical laws, they paint absurd situations, they generally make no sense.

Here's how it works.

You show a neural network a number of examples, the only requirement is that you're more often right than you are wrong (a far cry from admitting you're wrong, besides there are advantages, both social and practical from never admitting you're wrong, for example : you do well in the democrat party).

One thing you always see in artificial neural networks, but only in severely malfunctioning patients, is overdimensioning. Humans nearly always oversimplify. Artificial neural networks nearly always overdimension.

When you sleep, your brain cuts the inputs to your nervous system (by 99%, not 100%, so in cases of extreme need you'll wake up, but in the dark, relative quiet, under a blanket, you'll basically be oblivious to your surroundings with input signals weakened by 99%). But it DOES NOT cut the processing of the brain.

So what you see in your sleep is pure fantasy. Fortunately this is exactly what is wrong in your model. So you apply negative feedback learning on fantasy (NOT on wrong examples, that leads to overdimensioning), and run with that. You allow wrong datapoints, you just disallow fantasy. Because fantasy, that's damaging, and you always fantasize. Wrong datapoints don't throw off a simplistic model much.

M. Simon said...

I don't think so tom.

I have experience with a person who never allows the pain of error to influence that person's behavior and the dysfunction is unimaginable.

i.e. Error never reprograms the system.

You know the old saw - if you aren't making mistakes you aren't learning. What I wanted to say is that if you block the pain that mistakes cause you aren't learning either.