Neural Networks Revisited

QUOTE(dude @ Feb 5 2007, 08:36 PM)

In essence, the type of generalization you are referring to, is the basic building block of what we call science. Observe phenomena, formulate a theory that explains in general the behaviours that you have observed (and test it too!). But IMHO sometimes we need contrasts to move further in our quest for knowledge and I think this is very important.

Think of it this way, we were used to burning coal and oil to generate heat for our power plants. They worked on the principle of combustion... But when we discovered nuclear power, we had to change our viewpoint that apart from fossil fuels, nuclear energy can also be used to generate electricity. We could have formed a general opinion that combustion is the only feasible way to generate large amount of power in a viable manner. By thinking different, creativity is encouraged and people tend to think out of the box. Finding general patterns seems great (as it seems that we would have to work lesser for future discoveries, I suppose) but might lead to stagnation of knowledge.

Hey!
I think I have misconveyed myself, so I will try to make things a little more clear.

I'm not saying that all scientific work should be aimed at generalizing (although I do believe that that is what brings the quantum leaps). As you have said, work at lower levels is important too, more so because how things work at a lower level can often give invaluable hints to the overall larger order of their working.

What I was trying to say is made more clear through an example. Consider any novel groundbreaking system which makes some sort of decision based on a distance parameter between two vectors.

The system qualifies as genuine research and ought to be hailed as such. Once, the system has been specified however, then the choice of the distance parameter becomes trivial. Any half educated technical student knows that there are several measures (Hamming Distance, Euclidean distance, Mahalanobis Distance,Correlation...) and others can be invented as you go along.So that if you have spent 15 years just trying variations of all these distance on the same system, I say something is wrong. And for these things to be published in IEEE journals as articles of cutting edge research is greater folly still.

Pick up any volume of Technical transactions and you will find discussion on things analogous to the example I have given above(e.g. "Modification of the Karhunen-Loeve Transform using matrices of all image classes rather their correlation matrix") as if the aim of their research was the distance measure rather than the system (which in the previous case was Facial Recognition). I can't help but feel that these guys keep forgetting what they started out to do.

A final word on this-I am also not against working these 'distance parameters' if it can be shown that one of them is more fundamentally related to the system than another.e.g. In the case of Self-Organising Neural Networks, A feature based approach can aid in preliminary input differentiation but if we want to draw conclusions from input data, then an Information Theory based approach is required. In this case, the determining principle for self-organisation needs to be changed and I am all for it.

I hope I have made myself clearer.