Blogotariat

Oz Blog News Commentary

Can AI make you happier with clothes?

April 22, 2018 - 09:40 -- Admin

There is a great scene in A Series of Unfortunate Events (the Lemony Snicket books and now Netflix series) where a couple live in a fashionable penthouse suite but the elevator in the building is off limits because elevators are presently ‘out.’ So everyone, happily or not, has to ascend and descend many flights of stairs despite having a perfectly, functioning elevator. It is left to us to speculate why elevators might be ‘out’ but it doesn’t take long to think of a health or environmental explanation to give it a (slight) ring of plausibility. It is a great plot device — to shut off an elevator without it just being ‘out’ (of order) as it has been for years in The Big Bang Theory — but still get the peril and isolation associated with being high up on a building. We don’t know who decided elevators were ‘out’ but no one seems to care as they blindly follow its dictates.

In an article in Racked, Kyle Chaykar, considers what will happen if algorithms tell us what is ‘in’ or ‘out’ as well as how we, as individuals, should dress. I’m not the type to read Racked, but this discussion was fascinating and I recommend that you delve into it before going on to read what I have to say here.

Chaykar describes an emerging set of innovative uses of artificial intelligence (AI) whereby innovators suck up data of what ‘looks good’ or some other metric of fashion and use it to determine what you should wear. The appeal of this is obvious. Many of us find choosing an outfit very difficult, tense and potentially time consuming. Wouldn’t it be great to cut away all of that and have a computer just do it for you? After all, there are many people (myself included) who are quite happy to cede such decisions to another person (in my case, my spouse) or to just abandon all pretence and just wear the same thing every day. A case in point … our ‘social’ leaders:


Here, by the way, is me from 2008 and 2018:


Unlike the billioniares, somewhat heroically, it is the very same clothes for me! This is the luxury of being a tenured professor. Don’t believe me? Go look at one.

The promise of AI for fashion is the promise of some science fiction future. I suspect that does not mean the future as imagined by us … which apparently brings us velour …

Instead, the way current machine learning works is to use data to serve up suggestions. In our book Prediction Machines, we observe that what AI does well is prediction. In terms of deciding what to wear, that means selecting, from the available set of clothes, a prediction of what you should wear. To achieve that someone (or more likely many people) must have spent time looking at images of clothes and classifying them according to some criterion. One way that might happens is to identify clothes that appear in fashion magazines or are worn by celebrities and, from the set of clothes available to you, select which ones are the closest match.

“Wait a second,” I can hear you thinking, “What if I don’t want what celebrities are wearing? What if they look good on them but not on me? What if they are not comfortable? What about the weather? What about the city I live in? What about what I am going to do today?” Don’t worry, Chaykar hears you too and these are concerns. Moreover, much of that is far from insurmountable. So long as the right training data is available, the AI can take all of those things into account. It won’t happen immediately but, in principle, there is no reason why the AI can’t predict exactly what you want to wear.

The issue, as Chaykar opines is, would you want an AI to do this? Let’s dispense first with the idea that you want some sort of agency. As Chaykar documents: you don’t have that.

Roland Barthes noticed this arbitrariness in his 1960 essay Blue Is in Fashion This Year. Barthes scrutinizes a fragment of text from a fashion magazine — “blue is in fashion this year” — to see where its thesis, that a particular color is particularly tasteful right now, comes from. His conclusion is that it doesn’t come from anywhere: “We are not talking about a rigorous production of meaning: the link is neither obligatory nor sufficiently motivated.” Blue is not in fashion because it is particularly functional, nor is it symbolically linked to some wider economic or political reality; the statement has no semantic logic. Style, Barthes argues, is an inexplicable equation (a faulty algorithm).

You are being served up clothing options that are designed and dictated by those looking to sell you clothes. You are even choosing amongst those based on what your friends suggest and your family will put up with. If you feel you have real choice here, that is pretty much an illusion.

In the end, Chaykar is not sure where this leads. The end state is perhaps a set of clothing lines and choices that are certified ‘algorithm-free.’ But that too is an illusion.

The bigger picture, surely, is: can AI actually help here? For instance, if you are like Jobs or Zuckerberg, you don’t want to spend time on clothing decisions. They opted for a rule that satisfied this constraint without incurring any cost in fashion errors (if you don’t count wearing a uniform as an error). In that situation, if you want to do all of that and wear something different each day, the AI may well be able to do the job.

What about at the other end? Can AI find clothes that are affordable but, let’s face it, don’t look that way? This may be possible but it is going to require clothing makers deciding to listen to the recommendations of AI in determining what to produce. In this regard, the proof will be in the pudding. If pitted against their human fashion designer rivals, can the AI predict more accurately what will sell? My guess is that they can in some domains but not in others. But what share of our clothing will be algorithmically generated is hard to say.

The opportunity is huge. We currently buy from a set of clothes that is available. Most of what we see in stores are things that we do not, and will not, buy. It is rare to find things that work. I understand this problem intimately. When I find clothes that work, I buy multiple sets of them. Since I also know that clothes don’t stay available for long, I go to great lengths to do this. For instance, I have worn the same type of Ecco shoes for almost a decade. These shoes were no longer for sale as of 7 years ago so I scoured the Internet and bought up every last pair in my size. Each one lasts about a year. Soon I will have to wonder if leaving the house is ‘shoe worthy.’ (What goes on under the hood is the same thing and there I can highly recommend Bombas socks and MeUndies; may they last forever unlike my last undies provider in Australia that apparently could no longer survive on my demand alone).

This highlights another message from Prediction Machines: AI will be deployed where the degree of errors in decision-making is the highest. Clothing manufacturing has huge error rates. It is ripe for disruption.

The challenges in doing that are immense. It is one thing to say, look at available data and design clothes that people will want to buy. It is another thing to work out how to train a machine to do that. AIs learn by taking data and minimizing the error rate for their predictions. For the temperature, this means: how close is your forecast to reality? But for clothing decisions, that is not obvious. To be sure, you can minimise the difference between supply and demand, but, in actuality, the clothing decision is potentially too complex. For instance, people may not be comfortable buying things that err on the side of being different as much as erring on being the same. They may, in some situations, put up with a ridiculous degree of discomfort and inconvenience — and I am thinking of women’s fashion here (I mean why can’t women’s jeans have bigger pockets? It is 2018 for goodness sake!) — over looks. In crafting your algorithms, the machine doesn’t know how to make those trade-offs. Someone has to decide. It is part of what we call ‘judgment.’ Machines can’t do it. We need people.

And that is just to manufacture clothes. Actually, working out what algorithm will get your a Zuckerberg-Jobs style decision each morning is another matter. There we are just as personalised in our preferences as we are when buying clothes. The AI will have to learn those preferences and, moreover, how they might change in subtle ways. Suffice it to say, I am optimistic AI can solve my clothing problems but I believe that for younger folks, it is a long way off.

Imagine, however, that AI is up to the job. Will that be it? Can we be relieved of our clothing decisions and task it all to a machine?

When I told you of the Lemony Snicket plot where elevators were ‘out’ and people blindly followed fashion, I did not tell you of what ended up happening (spoilers ahead!). The villain in the book eventually uses the fact that people would follow dicta as to what is ‘in’ or ‘out’ to manipulate those people shamelessly to their own designs. This caused me to wonder: What if villains in our world did the same? While the scenarios of cyberattacks and hacking of the Internet of Things come from tampering with heating systems and ovens to causes fires, what if, would be terrorists hacked daily fashion algorithms? Think of being outside one day and seeing everyone dressed in the most ridiculous stuff. Or worse, just you or a small fraction of the population wearing something that looked crazy to everyone else? Wouldn’t you just die? In the end, it would not surprise me if that fear, that my personalised fashion algorithm started working against me, was ultimately what kept them being adopted.