Supercharging Deep Learning: Elevating Your Models and results via first principles

Saksham Gupta (~sgdescent)


0

Votes

Description:

Introduction:

Drawing from my experiences at top-tier organizations such as Microsoft Research, Zomato, and Boson AI, I've observed a recurring issue: many developers struggle to write effective deep-learning code. The process involves much more than just calling model.backward() and model.eval(). Often, mistakes in our pipelines manifest not as obvious bugs but as subpar results.

In my talk, I will address the importance of building robust, well-typed, and thoroughly tested ML pipelines. I will share insights and practical tips to help you enhance your Deep Learning applications, ensuring they are more resilient and efficient.

Objective:

This talk aims to identify common pitfalls in Deep Learning development, which are often overlooked due to the fast-paced nature of the field, and provide strategies to avoid them.

Join me to learn how to refine your Deep Learning code, improve your models, and ultimately make your Deep Learning applications "go brrrrr"!

P.S. This is not just a Horace He appreciation talk (well, maybe a little).

Prerequisites:

  1. Interest in Machine Learning and Deep Learning
    1. Previous experience writing DL code in any framework of choice!

Content URLs:

  • https://horace.io/brrr_intro.html
  • https://discuss.pytorch.org/t/debugging-cuda-out-of-memory/130342

and so much moreeee!

Speaker Info:

Saksham graduated from VIT Vellore in 2022. During his undergraduate studies, he focused on Machine Learning for education and conducted ML research at prestigious institutions like Carnegie Mellon University (CMU). He was also a MITACS Globalink scholar. Saksham interned and later worked full-time at Prodigal (YC S18) as part of the platform team.

To further his passion for AI, Saksham joined Microsoft Research Redmond as a pre-doctoral research fellow on the Program Synthesis team, where he worked on using large language models (LLMs) for code synthesis. Following this, he joined Boson AI as a research scientist, collaborating with renowned experts who have developed significant projects like VLLM and SGLang, including Alex Smola.

After a brief stint as a contractor at Boson AI, Saksham recently joined Zomato, where he is now dedicated to advancing AI initiatives across various projects.

Speaker Links:

  1. https://www.linkedin.com/in/higgsboson1209/
    1. https://scholar.google.com/citations?user=Me7NQeoAAAAJ&hl=en

Section: Artificial Intelligence and Machine Learning
Type: Talk
Target Audience: Advanced
Last Updated: