Building Visual Search Products at Pinterest
Pinterest is used by users as a source of inspiration to discover ideas for their creative outlets. To enable users to take action on these inspirations, Pinterest has a collection of products that a user can purchase. But there’s one part missing from this funnel from discovery to inspiration to action, a way to link these inspirational images to the products within the image. In this talk, I will discuss how Pinterest uses deep learning algorithms to power our real-time visual search system to enable this object discovery experience. We will share our approach for scaling this system to index billions of images while maintaining a latency of a fraction of a second and for incorporating user feedback to create a self improving system.
- How does a visual search backend architecture look?
- How to train embeddings for visual similarity
- Multi-task learning is useful to try
Andrew is a software engineer and currently leading the visual search project at Pinterest. He studied computer science at Berkeley (BS in CS) and Stanford (MS in CS). He is currently working extensively with the Berkeley Caffe team to scale up deep learning to the task of large-scale visual search.