KittenAuth

[the_eggwhite]

This amuses me.

I reckon I could write something that'd fool it (see below), but it would be processor intensive enough that the people with reasons to need to fool it probably wouldn't want to bother.

How to fool it:
a) MLP using spread-sample colourvalue inputs. Train with feline/non-feline images with correct classifications provided for desired outputs. When presented with new images would categorize into feline/non-feline results.
b) MLP trained using spread-sample colourvalue inputs, Train with feline images only, using input pattern for desired outputs. When presented with new images, feline images would produce an output similar to input, non feline would produce an output very different from input.
c) Kohonen using spread-sample colourvalue inputs trained with a sample of feline images. When presented with subsequent images, error vector would be of lower magnitude for feline images, higher magnitude for non feline.

There are probably easier or more appropriate ways, but these are the ones that leap to my mind.

Comments

[stealthracer.livejournal.com]

It is very interesting you should mention this. I've been thinking about picking up one of my old projects that involves image recognition and classifying. :-)

[eggwhite.livejournal.com]

Image recognition was never really my strong point, but novelty detection was... and there is some crossover if all you're trying to determine is "X" or "Not X".

Naturally, there'd need to be a chunk of preprocessing in front of my suggestions here, too. Also, it's likely that the "colourvalue" I mention would probably use luminosity or somesuch, rather than anything actually based on the specific colour - otherwise it'd learn to recognise colouration too much rather than shape or proportions.

[anonymous-james.livejournal.com]

Without clicking on the link, this makes even less sense than yesterday's post. Go figure.

[eggwhite.livejournal.com]

Yesterdays was just written in "Aiee! My Brain" language. This one actually makes sense if you're familiar with the field I'm talking about...