TUTORIAL – HALF DAY SESSION
Self-Supervised Learning: An Efficient Learning Technique to Learn Representations from Unlabeled Image Data

Introduction : Self-supervised learning methods are emerging rapidly in last few years. They have been used to solve several problems related to unlabeled data. Uses of these methods in fields like computer vision and natural language processing have shown many great results. In computer vision, deep learning is essential for in-depth and accurate analysis of images. Specifically, deep convolutional networks are most appropriate for extracting meaningful features from images. A huge amount of labeled data is required to train these deep convolutional networks for more precise predictions and better data usability. However, manual labeling of images is time-consuming and expensive for domain experts. In addition, the major issue with the manual labeling of the huge dataset is the bias among human annotators. Therefore, applied machine learning/deep learning is essential to process the dataset having largely unlabeled data.

Applied machine learning/deep learning includes unsupervised learning, semi-supervised learning, and Self-Supervised Learning (SSL) to learn from unlabeled data. Among them, SSL has been widely used in recent years to process image data due to its solid advantages over unsupervised learning and semi-supervised learning. It reduces data labeling cost and leverages the unlabeled data pool. SSL is a machine learning paradigm where a model, when fed with unstructured data as input, generates data labels automatically, which are further used in subsequent iterations as ground truths. The fundamental idea of self-supervised learning is to generate supervisory signals by making sense of the unlabeled data provided to it. It attempts to learn the visual representations of the data using proxy tasks perceived as pretext tasks. Pretext tasks are responsible for learning the prominent visual representations of data to use the learned representations or model weights obtained in the process for the downstream task. Today, self-supervised learning is most widely used for solving computer vision problems such as image classification, object detection, semantic segmentation, or instance segmentation.

The topic of proposal is emerging and demanding. It connects four contemporary areas of research: AI, Computer Vision, Data Science and Healthcare. This tutorial will provide attendees with a collective update on developments using self-supervised learning, challenges, opportunities and future research directions.

Expected Length of the Tutorial : 3 hours

Outline of the Tutorial :

• Labelled and Unlabeled Image Data for Computer Vision Applications (10 mins)
• Challenges in Processing a Large Volume of Unlabeled Image Data (10 mins)
• Various Machine Learning Techniques to Learn Representations from Unlabeled Data (10 mins)
• **Brief discussion on disadvantages of unsupervised learning and semi-supervised learning to process unlabeled image data. Subsequently, self-supervised learning will be discussed with its advantages over unsupervised and semi-supervised learnings
• Self-Supervised Learning to Learn Representations from Unlabeled Image Data (10 mins)
• Pipeline of Self-Supervised Learning (15 mins)
• Taxonomy of Self-Supervised Learning Methods (20 mins)
• Self-Supervised Learning for Computer Vision (30 mins)
• Our Proposed Novel Self-Supervised Learning Framework (20 mins)
• Use Case: Surgical Tool Detection from Laparoscopic Surgical Video (20 mins)
• ** Discussion of use case will include problem statement, the proposed solution, dataset, results and analysis, challenges faced, and
• Working demonstration of the use case (10 mins)
• Challenges & Future Research Directions (15 mins)
• Q & A (10 mins)

Learning Outcomes : After attending this tutorial, the conference attendees will

• Be familiar with different methods to process unlabeled data
• Get insights on how to process unlabeled image data to develop solutions to real life problems
• Receive conceptual understanding of self-supervised learning to process image data
• Get exposure to self-supervised learning-based use-cases.

Level of the Tutorial : Intermediate (assuming that attendees will have fundamental knowledge of machine learning)

Instructor :

Dr. Mayuri Mehta is a passionate learner, teacher and researcher. She is currently working as a Professor in the Department of Computer Engineering, Sarvajanik College of Engineering and Technology, Surat, Gujarat. She is also an International Relations and External Affairs Officer, Sarvajanik College of Engineering and Technology. She received a doctorate in Computer Engineering from National Institute of Technology (NIT), Surat. Her research interests include Data Science, Machine Learning/Deep Learning, Medical Image Analysis, Health Informatics, Computer Vision and Computer Algorithms.

Email : profmayurimehta@gmail.com , mayuri.mehta@scet.ac.in

Webpage : http://scet.ac.in/employee/prof-dr-mayuri-mehta/

LinkedIn : https://www.linkedin.com/in/dr-mayuri-mehta-a1484418/recent-activity/all/