Five years ago, if someone had tried to convince me to invest a penny in a product made without any human intervention, I would have laughed at them. But when I got to know that among many others, two of India's leading fashion e-commerce site’s brands are primarily made by AI, I realized it would be a missed investment opportunity. My one penny would have been worth a pretty penny today if I had invested in AI-driven fashion at the right time. While upset at my lack of foresight, I was also thrilled to know that among so many famous brands, these AI-driven brands are rated as customers' favorites. The interesting thing for me as an AI-ML professional (and potentially many of you), is to understand how artificial intelligence becomes smart enough to beat fashion designers with decades of experience and the creative instinct which has made them household names.
Let us Understand the Science behind it: -
The vanguard here is GAN (Generative Adversarial Network). A GAN is an ‘unsupervised’ learning model, which means that it does not try to estimate or predict like a traditional regression/classification model. Supervised models work well when there is something tangible that needs to be predicted and it can help us find out what factors affect the predictor variable. An unsupervised model on the other hand tries to look at a lot of data and identify the pattern behind it.
Thus, the GAN internalizes the attributes of a specific style by merely looking at a large number of examples, and it can subsequently apply that style to an existing piece of apparel. The process of internalization is driven by neural networks which have been discussed in detail elsewhere and are useful pattern recognition tools. What is more interesting is the structure of the GAN after the pattern is recognized.
The GAN is made up of two deep neural networks that work together to learn efficiently from raw input.
○ The Generator uses the pattern recognition algorithm to generate ‘fake’ or ‘new’ images to ‘fool’ the other network.
○ The Discriminator uses the pattern recognition algorithm to classify if an image is ‘real’ (and thus part of the original data) or ‘fake’ (which was created by the Generator)
○ The discriminator makes every effort to discriminate between actual and fraudulent photos. Both of them progress alternately, becoming increasingly adept at their respective roles.
○ If the Generator can fool the Discriminator often enough, you will effectively have a ‘new’ image which is close enough to the original images but is not real
The process of developing new designs always start with multimodal search. Here multimodal means where text and images can be used as a search query as shown in the image below. The input image consists of a red t-shirt and the text prompt says black colored and long sleeves. The GAN’s job then is to use this multimodal search to create a new t-shirt which will the characteristic of the original image along with the characteristics provided by the text prompt.
Working of Multimodal query with GAN
This is where the triplet loss function comes in which uses the anchor image, and a positive and negative distance vector is used to show that the distance between baseline input and positive input is lower than the distance between baseline input and negative input. Refer to the image below:
Training set using triplet loss
In this way, GAN helps in getting design patterns that are popular among customers, through pattern recognition, textual prompts and using triplet losses which can keep continually improving the algorithms. The only issue with GANs, like most neural nets, is the requirement of large datasets to train models for pattern recognition. But the world wide web provides enough data from competitor websites, fashion leaders or other social media platform to support this need. The AI algorithm develops a variety of combinations using a ton of data from customer records, bestsellers, and social media popularity before independently determining which design would be the "next bestseller." That implies that the computer is doing all the data analysis, design creation, and decision-making on its own!
Did we ever think that fashion would come with a tag like in mass produced food, "No Humans were Involved in the Processing of this Item”? Artificial intelligence seems to be becoming the real intelligence behind mass-produced fashion, yielding better results, and beating experienced fashion designers, as these super smart data models are fed data that belongs to millions of customers and thus helping in better and more accessible design.
Commentaires