Multi-task Deep Learning for Image Tagging

Recorded at Rev 2 | May 23-24, 2019 | New York Day 2 - Practitioner - SAS A fundamental characteristic of human learning is that we learn multiple pieces of information simultaneously. We can describe an image verbally because we are natural multi-task agents. A comparable concept in machine learning is called multi-task learning (MTL) and it has become increasingly useful in practice. A common MTL use case is image tagging. For example, a retailer can use MTL to identify visual attributes for clothing items. Multiple attributes are learned simultaneously such as the type of clothing, texture, color, pattern, gender, and fit type. The tagged results can be used for customer profile analysis to make purchase recommendations. With a set of personal photos, it is possible to infer the fashion style of the shopper by analyzing the attributes of clothes and then recommend other clothing items for purchase. Tagging can also be used for retrieval systems like image search, or as part of feature engineering. In this presentation, we build a multi-task deep learning model using DLPy to tag fashion clothing items. Convolutional neural networks show extraordinary performance for image classification and object recognition applications. DLPy is a high-level and easy-to-use Python API for SAS Deep Learning models. We explain how DLPy can be applied to data preparation, data processing, multi-task model building, assessment and deployment for image tagging.