Diagnosing Your Model: Learning How To Debug Deep Learning Models Using Visualization For Medical Images  

Building your models isn’t tough, it’s debugging models that is tough – especially ones that fail silently. How to understand that the model is working in failsafe mode – is high performance metric like AUC enough ? In this talk, we discuss why visualisation is an absolute must-have in your toolbox and discuss some case studies about how we understood our the working of models better using visualisation. Apart from that, we would be discussing all methodologies till state-of-the-art techniques used for visualisation of ML & DL models.

The talk consists of the following segments:
  • Importance of visualisation
    • Especially for medical imaging
  • Case studies of how we debugged some problems with use of visualisation
  • Overview of visualisation methods – perturbation & gradient based
  • Perturbation
    • Occlusion + LIME +   Integrated Gradients
  • Gradient Based
    • GradVis + GradCAM + LRP + DeepLIFT + Shapley
  • Libraries you can leverage to get started


Rohit Ghosh

One of the Top Data Science Influencers in India (Digital Vidya). Winner of Zinnov Award 2017 – Technical Role Model – Emerging Technologies (Senior Level). Joy Mustafi is Director and Principal Researcher at Salesforce, primarily responsible for leading the Salesforce Einstein and other intelligent cloud platform in India. Joy has overall sixteen years of experience in corporate, research and academic world.

Rohit is a Founding Member & AI Researcher at Qure.ai with 10+ publications in multiple medical journals and global radiology conferences, including publications in The Lancet. Most of his research at Qure is centered around detecting critical abnormalities from CT scans of the head and brain. He also focuses on research methodologies involving interpretation (unboxing black box) of deep neural networks for computer vision.

He also teaches data science at multiple academies and campuses. In addition to computer vision, he also keenly follows developments in reinforcement learning (mostly related to trading strategies) and writes about papers related to data science and ML- based trading strategies here – https://rohitghosh.github.io/.

He is a 2014 IIT-Bombay graduate and an ex-investment banking professional with Nomura.

Buy Ticket
Social media & sharing icons powered by UltimatelySocial