Deep Learning for Computer Vision Applications: A Case Study

Tutorial, HiPC 2019 - December 17, 8:30 am to 1 pm

Abstract: This hands-on tutorial introduces a deep learning framework for the problem of automating inventory management using computer vision methods. The motivating business problem is to automatically recognize the number and type of objects in a warehouse and provide real-time status updates to the owners. Our tutorial introduces a pipeline for this computational challenge, and this pipeline could potentially be applied to other image recognition problems and different application domains. The solution uses the Tensorflow architecture and the LabelImg image annotation tool, and can be extended for the task of semi-automated labeling. 

The tutorial discusses several topics related to high-performance computing, including reducing model training time, effectively using compute resources, scalability issues in training convolutional neural networks and recurrent neural networks, and pipeline performance monitoring. We will also illustrate how real-time feedback from model predictions can be incorporated. The robust infrastructure is built using Angular, MongoDB, and Python. Our solution aims to avoid inventory waste and under- and over-utilization in offices and warehouses, thereby reducing operating costs. 

This tutorial targets attendees of all skill levels, including students, researchers, and data scientists in the IT industry. A basic knowledge of machine learning algorithms and Python are assumed. Tutorial attendees are expected to bring their laptop computers. We will use Jupyter notebooks for the hands-on exercises.  

 

Schedule 

Problem overview, challenges and high-level approach – 20 minutes 

Solution overview, technology stack used – 20 minutes 

Hands-on Session 1 

Installation/setup, basic demos (image processing and detection) – 60 minutes 

Demo walkthrough 

Performance tuning for machine learning: Overview – 20 minutes 

Model selection, hyperparameter tuning, scalability – 20 minutes 

 

Hands-on Session 2 

Installation/setup, trying different models, tuning libraries – 60 minute 

 

Presenters 

Anuradha Karuppasamy is a Senior Data Scientist at Ericsson’s Global Artificial Intelligence Accelerator (GAIA) in Bengaluru, India. Her areas of expertise include deep learning, computer vision, biometrics, and natural language processing. 

Anant Gupta is a Senior Data Scientist at Ericsson GAIA. He was previously with Morgan Stanley and worked on client analytics, machine learning, and data engineering-related projects. 

Mahesh Babu Jayaraman is a Principal Engineer at Ericsson GAIA and works on machine learning and AI methods for telecom network management decision support, analytics, and automation. His areas of specialization include full-stack research and development in network and element management systems, telecom operations support solutions, network planning, and policy management. 

Scroll Up