Computer geeks are just as interested in fashionable clothing as anyone else, though they tend to apply the mental hammers they possess to driving the nail. There are any number of projects that are applying contemporary AI to the alleged problems of finding (and creating) fashionable garments and outfits.
Whatever the problem may really be, this research essentially treats it as a “recommender” system, a la online shopping and streaming services. This leads to two products, a “virtual stylist” to advise you on what to wear, and a “trend spotter” to advise producers about what to sell.
So who would want to use an AI recommender?
This kind of technology can probably detect group uniforms pretty easily, especially if social media and other metadata are included in the data. This may be useful for market intelligence, but probably not terribly useful to individuals. If you have to have a computer to tell you how to dress like the people you admire, you’re pretty lost.
On the other hand, some people might enjoy having a “virtual stylist” who helps them construct and maintain an individual look. For that matter, the ability to generate something that is in the style of X, but new, would be exactly what you might be looking for. What will Susie be wearing today? She’s always ahead of the curve. Etc.
Underlying these ideas are collections of data about what people are wearing, and the usual “people who liked this, also liked this other thing”. Grist to this mill are images from the internet, social media posts, personal shopping history, and metadata about who’s who and what they do and want. Many readers will recognize that this is also the data and technology used by advertising and intelligence services, who are looking to predict specific kinds of individual behaviors.
This fall, a team from UCSD and Adobe report on yet another permutation of this technique, which uses sophisticated image processing and machine learning to create “look plausible yet substantially different from” the examples . (Note: their paper cites quite a bit of earlier work, which is worth looking over.)
The most interesting idea is to use the “user X image” preference data to generate new items that are predicted to be attractive to the user.
“a richer form of recommendation might consist of guiding users and designers by helping them to explore the space of potential fashion images and styles.”
This is pretty neat, and most machine learning approaches can’t do it nearly as well as what this group has done.
The technical details are non-trivial, see the paper.
This work is interesting, but there are a number of questions raised by this work.
First of all, it’s far from clear that this is addressing a problem that anyone needs to solve. For those who go beyond pure utilitarianism, “fashion” is a signaling system. What you wear is supposed to send messages about yourself.
The two messages most commonly sent are either “look at me, desire me” or “I belong to this group”. Note that these are somewhat contradictory messages, asserting either individuality or conformity, respectively.
How do these messages relate to “preferences”?
Second, the authors suggest that this technology could be used as an aide to designers:
“In the future, we believe this opens up a promising line of work in using recommender systems for design.”
“We believe that such frameworks can lead to richer forms of recommendation, where content recommendation and content generation are more closely linked”
In other words, this is a feedback loop, from design to user’s reception, back to designer.
As I have said before, this would seem to be a mechanism that pushes designers to produce “more of the same”; scarcely a formula for creativity. But perhaps the AI is actually chasing user’s reactions which will eventually reject “more of the same”, it is retrospective—not a formula for being a fashion leader.
On the other hand, this technology does generate new designs, though it is hard to judge just how creative it is. Also, the technique learns continuously, which seems critical to me-preferences change. Done right, the AI model might be a good way to represent a target user and to generate designs that hone in on their (momentary) preferences.
However, there are lots of underlying issues.
To the degree that a person wants to make fashion statements, this process of digging out “preferences” and generating examples that exemplify them is only indirectly related to whatever the intended statement might be. The AI knows little, if anything, about the semantics of the clothing in the images, which are highly subjective in any case.
Many fashion preferences are based on social aspirations, such as the desire to emulate a celebrity and / or fit in with a clique. These factors are not only not visible in the image, they are completely missing from the concept of a personalized design. Don’t you want a social design?
The entire process is based on (small) 2D images with standard poses. This is a very impoverished set of information. Users have been trained to deal with tiny 2D mages by the internet, but this is not a full representation of “fashion” or anything else. Clothing is 3D and bodies move, and both exist in physical context (e.g., a dance floor). These factors are missing from both the input and output of this AI.
The training uses ratings and other inputs from observers as proxies for “preferences”. It isn’t clear exactly what these data actually represent, and there is a very real possibility that there are multiple “communities” with different preferences all using the same online services. Averaging across multiple sub-cultures will produce a meaningless common denominator. (The research project used tiny, probably homogeneous samples, so it doesn’t explore this problem.)
Building a tight feedback loop between user preferences and the AI opens the possibility of hacking or gaming the system. Flooding the inputs with ratings and supposed positive comments could manipulate the recommendations. I could easily imagine a PR campaign that combined celebrity product placement with ‘AI placement’ that biases the results in favor of the product.
Worse, there could well be AIs gaming other AIs. It could devolve to the point where everyone has a ‘virtual fashion advisor’, and all the AIs are tracking the behavior of all the other AIs. Kind of like professional fashionistas do. Sigh.
I have to ask, just who is the AI serving? The researchers seem to believe that the designer’s interests are aligned with the user, but I don’t think it is that simple. Designers usually are or work for producers, who aim to sell the product. Is the recommender helping the consumer or promoting consumption?
We all know the answer to that question. Technology just like this is already used by advertising and intelligence to track and predict the behavior of individuals. This behavior modelling is not for the benefit of the subject, it is for the benefit of the wealthy and powerful.
If these types of system come into wide use, it will probably change concepts of fashion recommendations. An on-line suggestion that “people like you, also liked these items” is not valued as much as a recommendation from a close friend. Similarly, trends spotted (or created) by AI will be considered second rate, compared to actual human trend spotting and creativity. People will work hard to outguess the AI.
I predict that the more virtual assistants there are, the more valuable a human assistant will become!
- Wang-Cheng Kang, Chen Fang, Zhaowen Wang, and Julian McAuley, Visually-Aware Fashion Recommendation and Design with Generative Image Models. arXiv, 2017. https://arxiv.org/abs/1711.02231
- Will Knight, Amazon Has Developed an AI Fashion Designer: The retail giant is taking a characteristically algorithmic approach to fashion. MIT Technology Review online.august 24 2017, https://www.technologyreview.com/s/608668/amazon-has-developed-an-ai-fashion-designer/
- Jackie Snow This AI Learns Your Fashion Sense and Invents Your Next Outfit: A new kind of AI system could create personalized clothing based on a shopper’s taste. MIT Technology Review online.November 16 2017, https://www.technologyreview.com/s/609469/this-ai-learns-your-fashion-sense-and-invents-your-next-outfit/