PyData Seattle 2023

Deep Learning Model Interpretability for Computer Vision based Models
04-28, 10:15–11:00 (America/Los_Angeles), Hood

Applied Deep Learning to computer vision has become very popular in the last decade. Many real-world problems related to detection and recognition are being solved by using popular open-source models. Many problems are very specific and off-the-shelf models do not work as it is. These models have to be trained with custom data to perform specific tasks. While training these models apart from empirical information related to training performance, there s no way to interpret the results from the deep learning models. In this talk, I will talk about various ways that we can use to interpret results visually for deep learning models.

This talk addresses the main problem of interpreting the results from deep learning models in computer vision, highlighting the importance of understanding how the model makes predictions. The talk will introduce various visualization techniques, such as feature maps and saliency maps, which can be used to interpret the results of the model. Furthermore, the talk will delve into the details of these visualization techniques, discussing how they can be used to understand what the model is learning and which parts of the input image are most important for making a prediction. Additionally, the talk will discuss the use of heuristics to interpret the results of deep learning models, which can also be used to debug models after training and improve performance metrics.

This talk is aimed at audiences who have a basic to intermediate understanding of Convolutional Neural Networks. By the end of this talk, attendees will have a deeper understanding of interpretability techniques and their application in computer vision and will be able to use them to improve model performance. With this knowledge, attendees will be able to make data-driven decisions and develop more accurate and reliable computer vision models.

Prior Knowledge Expected

Previous knowledge expected

Sumedh is a Senior Machine Learning Engineer with more than 6 years of work experience in the field of Deep Learning, Machine Learning, and Software Engineering. He has a proven track record of single-handedly delivering end-to-end engineering solutions to real-world problems. He works at the intersection of engineering, research, and product and developed Deep Learning based products from scratch that’s been used by thousands of end customers. Currently, Sumedh works in R&D where he works on Applied Deep Learning with fewer data and has several granted patents and several more applied. Sumedh studied master’s in computer science focused on AI.