NCSA’s Center for AI Innovation (CAII) in collaboration with NCSA’s Delta/DeltaAI team are offering a Practical Machine Learning Training Seminar Series during the Spring 2025 semester. One of NCSA’s core goals is to provide compute resources to researchers nationwide and the expertise and training necessary to fully utilize these resources.
The seminar will be taught by Priyam Mazumdar, a PhD student in Electrical and Computer Engineering and a researcher at the National Center for Supercomputing Applications (NCSA) at the University of Illinois.
This 10-week training program will cover topics on building and training machine learning (ML) models, ranging from beginner level to complex deep learning (DL) models trained on multi-GPU nodes of the Delta supercomputer. This seminar is open to all, including graduate students, undergraduates, and particularly domain scientists whose primary affiliation is not with a STEM program or department. This seminar aims to equip participants with tools and skills for independent research.
Last semester’s seminars focused on a wide range of topics for vision and language tasks, but this semester we will focus on auto-encoders as a technique to solve one of the most frequent problems in ML: the curse of dimensionality. Most lessons will be a mix of live-coding and light derivations to provide the necessary theoretical background. By the end of the series, participants will be able to implement these models from scratch with little use of packages other than PyTorch.
Registration Information
To attend the seminars, please complete the registration form that relates to your preferred type of learning environment (In-person or Virtual). Participation will be on a first-come, first-served basis and limited to:
Those who complete the series will receive a digital badge which can be used in social media or portfolios to showcase skills and achievements. The lessons will be taught simultaneously in-person at NCSA and on-line via zoom.
What you will learn
This git repository contains all the relevant training materials.
Prerequisites
VERY FEW! Again, the purpose of this series is to take someone who knows little about this field and teach them to re-implement/train a set of State-of-the-Art architectures. If you have seen a little bit of Python and have some motivation you can do this!
Our teaching philosophy
Advanced mathematics can be a barrier to learning and initially be intimidating for people just beginning their studies in Machine Learning. We want to clarify that as this course progresses some references will be made to mathematics, they will be a very applied and practical application of neural networks rather than a theoretical endeavor.
Before working on a PhD in ECE, Mazumdar,s background was in Neuroscience, so he understands the thought process and learning style differences between engineering and non-engineering students. The aim of this series is to teach as intuitively as possible and (hopefully) make it exciting! This also makes the classes different from the ML/DL/AI courses offered by University professors. This series acts as a bridge connecting the theory learned in those classes to practical implementations for research. There will only be one day of some heavier math, when the variational autoencoder loss function is implemented.
What you need to bring
The first 2/3rds of the seminar will be done mostly on Google Collab (we will discuss what this is on the first day, so everyone will have a similar basic knowledge). After that we will move over to the Delta supercomputer and walk everyone through gaining access. So, bring a laptop, have a google account, and you should be good to go.
Course format
Every session will focus on live coding. We will build in front of you so you can see exactly how it all works. We do recommend following along with the instructor, so everything makes sense.