Presented at virtual Activate 2020. ml4ir is an open source library for unified training and serving of deep learning models for search relevance. Built on top of tensorflow 2.0+, ml4ir is designed for scale using TFRecord data pipelines. ml4ir is built as a network of loosely coupled deep learning subcomponents. This allows users to define custom sub-models and combine them using a simple pluggable interface to build really complex models for a wide variety of applications. Alternatively, ml4ir can be used with little to no code only interacting via configuration files. This architecture allows ml4ir to be an ideal place for search data science collaboration with users being able to share different neural network layers with each other. ml4ir models are being used in production environments at Salesforce today as they are compatible with tf serving and also come packaged with the necessary code to deploy to JVM based environments.
Jake Mannix, Search Relevance Architect, Salesforce.com, Inc.
Ashish Bharadwaj Srinivasa, Senior Data Scientist, Salesforce.com, Inc.
Data Scientists and ML Engineers in the industry who want to learn about a component based deep learning abstraction for building enterprise ready models for search. The session would also be of interest to the Learning to Rank community and anyone interested in building deep learning models even with limited python experience.
The audience will learn about a new open source deep learning library for search. They will learn how to configure and customize ml4ir to train and serve a Learning to Rank model. This will enable them to onboard their applications and training data to ml4ir to build production ready models.