Multitask Learning For Image Tagging Wayne Thompson SAS

Multi-task Learning For Image Tagging Wayne Thompson, SAS @Thompson_Wayne (Twitter) Copyright © SAS Institute Inc. All rights reserved.

Agenda Machine Learning with SAS & Domino Data Lab SAS Deep Learning Toolkit Multi-Task Deep Learning Copyright © SAS Institute Inc. All rights reserved. Demo & Closing Statements

Deep Learning More Compute Power Larger Data Sets More Complex Models Copyright © SAS Institute Inc. All rights reserved. Better Algorithms Deep Learning

Sample Deep Learning Use Cases @SAS Person/Object Identification Three-Dimensional Scans 1 2 3 4 Traffic Surveillance Insurance Claims Copyright © SAS Institute Inc. All rights reserved.

SAS DL Compute Platform ESPPy Copyright © SAS Institute Inc. All rights reserved.

Multitask Learning Hard Parameter Sharing Soft Parameter Sharing Company Confident ial – For Internal U se O nly Copyright © SAS Inst itute Inc. All rig hts reserved.

Primary MTL Advantages 1 Implicit Data Augmentation 2 3 Attention Focusing Eaves dropping Company Confident ial – For Internal U se O nly Copyright © SAS Inst itute Inc. All rig hts reserved.

Candidate Applications Banking Fraud Detection Government Heath and Life Sciences Manufacturing and Energy Retail Medical Image Analysis Defect Detection Image Tagging Smart Cities Identification Political Campaigns Company Confident ial – For Internal U se O nly Copyright © SAS Inst itute Inc. All rig hts reserved.

Fashion Items Tag Attributes & Recommend Items • Currently have separate models to predict • Color • Type of Clothing • Texture • Build one multi-task deep learning model to predict all labels. Company Confident ial – For Internal U se O nly Copyright © SAS Inst itute Inc. All rig hts reserved.

Demo Path Model Training and Edge Deployment with Viya SAS® VDMML SAS® ESP Load Images Raw Video / Image Data Ingest ASTORE Load Images Process Images Video Footage/ Streaming Images Preprocess Images Train Deep Learning Models Score Images In-stream Generate ASTORE Post-process Computation Company Confident ial – For Internal U se O nly Copyright © SAS Inst itute Inc. All rig hts reserved. Feedback into Control Systems for Optimization and Triggering Alerts

Demo Path Enrich Store Train ETL Score Automated Response Copyright © SAS Institute Inc. All rights reserved.

SAS DLPy & ESPPy Demo Copyright © SAS Institute Inc. All rights reserved.

Thank you. Copyright © SAS Institute Inc. All rights reserved.

Notes • Challenges - viewpoint variation, scale variation, intra-class variation, image deformation, image occlusion, illumination conditions, background clutter Copyright © SAS Institute Inc. All rights reserved.