[LOCKSUIT]

Ok korrelan n all time to solve something here..

ANNs use backprop to generate weights in a hierarchy, ahem, representations, mathematical representations, abstracting the data more and more farther the deeper you go up the layers until the data deep in is unrecognizable and is so mathy, representing something deep, RIGHT?

Now, drumroll,

why mathematical representations? Or IOW why representations?

I mean, after deploying the trained net, the representations help see the parts of a fed in image of a cat and output the correct class 'cat', right?

Who says you need to make some numeric representation in the next layer after the data input layer? Why?



Also teach me some key facts about Representations in ANNs.

[ranch vermin]

What your representing is more important than how you represent it,  and how you explain it is completely personal to the scientist at hand,  so expect great conflicts when going to learn things off other people.  Its life.

[Art]

Much akin to an ordinary video game. Characters/objects move around, limited by their "world".

What we don't see is the underlying math that represents said object and determines how, when and where it is allowed to move and at what speed.
Thousands of calculations are performed in the background, formulas tested, measured and exacted. All of those If...Then scenarios carried out and if certain conditions are met, you might win that level and on to the next.

Without the math, formulas, equations, etc., all you've got is hardware...same with computers, robotics, and other electronic devices.

[Korrelan]

I’m not quite sure what you’re asking but…

abstracting the data more and more farther the deeper you go up the layers until the data deep in is unrecognizable and is so mathy, representing something deep, RIGHT?

Erm… yes and no.  Let’s use vision as an example. 

A deep convolutional network starts by looking for very small abstract features, bits of lines, parts of circles, gradients, etc. At each successive layer these features are combined to create ever more recognisable larger features.  So… technically the data, actually gets more recognisable as it moves through the schema.

The ambiguity of the data to a human viewer is simply because it’s a computer algorithm doing the selection/ matching process.  The computer is not using a human compatible/ equivalent methodology, it’s a mathematical approximation.  The computer is selecting and combining features ‘it’ finds relevant, based on the algorithms being implemented… these selections are usually not compatible with our visual/ sensory systems.

Who says you need to make some numeric representation in the next layer after the data input layer? Why?

That’s just the way it’s designed to work lol. The whole schema/ design is based on this method.  It’s a computational approximation of a human system. The numeric representations are the computers equivalent of bits of an image; an eye, a nose, etc.  These are what are combined to make a more detailed representation in the next layer.

The whole idea of a NN is to combine/ filter/ match the sparse/ complex input data into bigger more recognisable chunks/ concepts as the data moves through the network.

 

[LOCKSUIT]

The deeper you go, the clearer AND more abstract it gets.....in ANNs!

I'm not arguing that how they represent features is bad. I'm arguing that how they represent features is how they in turn decide what to represent, and that is bad!

Why wouldn't you want to store man AND woman AND human? Why learn the MATHEMATICAL REPRESENTATION "human" and drop the data man and woman and human!? Is it because mainstream backprop needs continuous variables? That forces it to use numbers. But continuous variables and this sort of backprop is only trying to figure out which connections to strengthen of say 5 connected to a feature. Right?