[ExI] Fashion

[Anders Sandberg]

OK, I might be totally lost every time I enter a clothes store or look 
at my wardrobe, but at least theoretical fashion design sounds like fun.

Fashion is defined as “the cultural construction of the embodied 
identity” by the journal Fashion Theory: The Journal of Dress, Body & 
Culture; it is not just about high fashion worn by models but just as 
much what Linux t-shirts to wear at hacker parties.

In this thread I guess we are looking at prevailing styles of dress, and 
especially how they change. It turns out that there are plenty of 
theories of how this actually happens and why (trickle down, sideways 
and up, status expression, cultural signals, markets etc.) A simple 
model is just that the current dominant style emerges somehow, gets 
copied by more and more people, and then either disappears because it is 
getting too common (not a good status symbol any more), market push 
(designers want to sell more stuff) or just because it gets stale 
(people like new things).

I have actually done models that act like that in my neural networks: 
consider a network of features where the current state of activation 
represents the current fashion. Connection weights represent what goes 
well together (based on visual contrast, bigger cultural values etc). A 
state of fashion is an attractor state that self-stabilizes. But as time 
goes on it weakens: in my network I subtracted a second weight from the 
connection weights, updating this second weight by slow learning. So 
after a while the fashion state jumps to a new attractor - often rather 
different, but influenced by the previous state (since that was where it 
started and by being significantly different). The result is a pretty 
complex dynamics that jumps chaotically between attractors. This seems 
to be generic: http://arxiv.org/abs/1208.0576 describes a model of 
fashions in names that has the same properties. 
http://arxiv.org/abs/cond-mat/0212267v1 has another model which is agent 
based, and shows self-organized criticality: big and small trends mix 
randomly.

Now these considerations suggests to me that predicting fashions 
automatically is not going to work, simply because the transitions are 
chaotic and noisy, and might even contain anti-learning: if they are 
predictable, people will shy away from them and do something else. But I 
don't see any problem with using methods like this to generate new 
potential fashions: train your software on past fashions (I would not 
use bitmaps, but rather some sort of feature description language), and 
then set it loose to suggest new possibilities that are "fresh".


To make this more transhumanist: here Ian Pearson is talking the future 
of fashion: 
http://www.podkolinski.co.uk/wordpress/wp-content/uploads/2012/01/fashion2.pdf

-- 
Anders Sandberg,
Future of Humanity Institute
Oxford Martin School
Faculty of Philosophy
Oxford University