Notes from Kyunghyun Cho’s lecture during Montreal’s Deep Learning Summer School
Recently, there has been a lot of talk around generative models as the next big thing in deep learning. There are articles about general adversarial networks, variational autoencoders, PixelRNNs, and perhaps some reinforcement learning augmentation. But why all this sudden interest? Why not pursue memory networks or other ideas? I am by no means an expert, but my suspicion is that generative models offer superior context and speed.
Although not intractable, the problem of training recommendation systems includes many difficult components that must be resolved in order to have practical use in the real world. Based on my current understanding, there are three separate problem domains:
Within the Beyond Clothing Ontologies: Modeling Fashion with Subjective Influence Networks, the authors Kurt Bollacker, Natalia Díaz-Rodríguez, Xian Li (from Stitch Fix) have presented a supplementary fashion ontology for modeling out the relationships of clothing that go beyond other existing ontologies by taking into account deeper subjective measures, namely influence networks, rather than just objective measures of clothing, such as length of sleeve or material. While related schemas on have been proposed in the past, they fall short from being useful because they fail to incorporate many subjective aspects of fashion, such as where a style originated from or how a particular design was influenced by past trends.
Given the goal of creating an optimal recommendation system, one could consider using a neural turing machine (NTM) with a “programmable” head performing read/write operations on the various inputs in order to generate a prediction. The inputs would be encoding of the situation (shopping for clothes), encodings of the options (sweater, jacket, T-shirt, dress shirt) along with their features of those options (blue/green, soft, polyester, wool), tons of other previously collected data points, and obviously a vector embedding of the individual for whom we are making the recommendation. The predictions would just the result of softmax classifier with thousands of possibilities, each one representing a cluster of clothing items. (There is no need to recommend just one item because eve a mobile UI can display 7-10 items fairly easily to let the customer make the final decision.)
