Augmenting Solr’s NLP Capabilities with Deep-Learning Features to Match Images
Matching images with human-like accuracy is typically extremely expensive. A lot of GPU resources and training data are required for the deep-learning model to perform image-matching. While GPU is something that most companies can afford, training data is hard to obtain.
At DataWeave, we crawl millions of products listed across e-commerce websites, and match them to deliver competitive insights to our clients. In the fashion vertical, however, text matching alone is insufficient to accurately match products, as product descriptions are usually not detailed enough.
We asked ourselves, is there any way of complementing information from product descriptions and titles to improve the accuracy of image-matching?
Solr is a popular text search engine known for its NLP capabilities. This talk will present an innovative way of storing deep-learning features in Solr, and augmenting Solr’s NLP capabilities to achieve elevated levels of accuracy in our product matching efforts.
1) Searching similar and exact images using deep learning (Importance and problems associated)
2) Solr – a popular text search engine
3) Augmenting Solr with Deep learning features
4) Self-taught hashing
5) Performance metrics
I work as a data engineer at DataWeave, a company that provides Competitive Intelligence as a Service for retailers and consumer brands. Here, I helped develop deep learning and machine-learning infrastructure for large scale product matching capabilities.
I am a keen enthusiast of open source projects, and have been closely associated with a project that integrated TensorFlow with DeepDetect.
I was among the top-5 finalists in the Xerox Research Innovation Challenge - 2016, and winner of the Jaipur Hackathon -2015. One of my projects - sign language converter (SLC) - was among the semi-final entries at TI Innovation Challenge India Design Contest 2015.
I have also co-authored publications that have been accepted in Applied Intelligence, Knowledge Based System, and International Conference of Machine-Learning and Cybernetics.