How to Train your Models with less data?

Syed Moinudeen (~moinudeen)


Description:

In Applied Machine Learning, Supervised learning is the most successful and widely used paradigm. For any task that can be solved with Supervised learning, we need a large dataset with Input-Labels pair. In real world applications, we often encounter scarcity of labelled data for training models in a supervised fashion. We usually have access to huge amounts of unlabelled data but the cost of labelling is too high or sometimes we don’t even have data points to label for the use case.

Over the past few decades, Machine learning researchers have been trying to find the golden equation that will solve this problem. To find an algorithm that will just learn from unstructured / real world data without any labelling effort being required. In reality the world is still far away from that milestone. However, We have made significant progress towards that goal by incorporating techniques that require relatively less data and/or compute to train models. This talk will cover some of these techniques like Transfer Learning , Self Supervised Learning, Semi Supervised learning, etc.

In real world applications, Transfer Learning has worked well in scenarios where the problems mentioned above were prevalent. In this session, we will discuss in detail about how Transfer Learning works for different problems we encounter in Computer Vision applications. We will see how this approach to supervised learning improves performance as well as the efficiency of the learning process. We will also take a look at how this approach is being used to develop new areas of ML research like self-supervised learning and semi-supervised learning.

Agenda:

  • Intro to Machine Learning in Visual recognition.

    • CNNs and Visual Features.
    • The problems with heavily Supervised learning.
  • Enter Transfer Learning!

    • Real World Scenarios for Applying Transfer Learning.
    • Performance and Cost Advantages.
    • Disadvantages.. Can we do better?
  • Automatic Labelling: Self Supervised Learning

    • Intro
    • How to apply self supervised learning to a task
    • Results
  • Semi Supervised Learning

    • Intro

Who is this talk for:

  • Deep Learning Practitioners

  • Students/ Researchers.

Prerequisites:

  • ML / Predictive system basics
  • ANN / CNN basics

Content URLs:

WIP ...

Speaker Info:

Syed Moinudeen is a Free software enthusiast and a Machine Learning Tinkerer. Currently working as a ML Engineer at Mad Street Den, dabbling with ML / Deep learning Algorithms for solving interesting problems in retail. He has also Volunteered for organizations like TN Artificial Intelligence Guild ( TN-AI Guild), Free Software Foundation Tamil Nadu (FSFTN) and others.

Speaker Links:

Github

Id: 1419
Section: Data Science, Machine Learning and AI
Type: Talks
Target Audience: Intermediate
Last Updated: