Abstract: Two related but separate topics pertaining to e-commerce will be covered: 1. Deep learning based visual recommendation 2. Deep learning based visual semantic embedding.

1. Deep learning based visual recommendation
Recommending catalog items that are visually similar to a catalog item the user is  browsing is an important problem in e-commerce. We refer to this problem as "CIVR" (Catalog Image based Visual Recommendation). CIVR is a very challenging task in its own right, due to the extreme variety among the catalog items that could be deemed similar - dress items maybe hanging or laid flat on a table  or worn by different models or mannequins having different complexions and/or hair color, standing in different poses etc. Furthermore, the human notion of similarity is extremely abstract and complex. Two t-shirts, one with a batman print and another with a superman print, maybe called "similar" by human beings, while, in terms of pixels, there maybe very little in common between the two images.

The main contribution of this paper is a deep CNN architecture for visual recommendation system,  which has been launched with much user satisfaction.  Our deep network generates an embedding vector - Euclidean distance between embedding vectors measures the (dis)similarity between images. Our embedding captures high level abstractions as well as low level details -  both of which are important for visual similarity. In this, we provide practical evidence based support in favor of the parallel deep and shallow network paradigm of Deep Ranking. The embedding generated by our network is  quite robust to background, pose variations, partial views etc. Finally, we provide experimental results on multiple related approaches to empirically justify our approach

2. Deep learning based visual semantic embedding

A multimodal embedding is learnt whereby product images and visual keywords describing them are jointly embedded into the same metric space. This is used to improve search results and to identify discrepancies between descriptive phrases and images of products.

Speaker Bio: Krishnendu Chaudhury is a Principal Data Scientist at Flipkart, Bangalore. He completed his doctorate in computer science, specializing in computer vision, at the University of Kentucky and has authored several publications in IEEE and ACM journals/conferences and has over a dozen patents in imaging and computer related technologies https://sites.google.com/site/krishhomepage/Home. Prior to joining Flipkart in 2015, he worked at Google, Mountain View for 10 years and Adobe Systems before that.