Scalable and Robust (Embedded) Machine Learning (SREML)

Course Overview

This course is concerned with the intersection of machine learning and HW systems. While machine learning (ML) is a ubiquitous task for various applications and demonstrates an outstanding quality in its results, it comes at huge computational costs. The course’s name stems from three different aspects:

1. We consider scalability aspects with regard to size of the neural architecture (larger is better in terms of quality) as well as the size of the computing system (larger is bad but required)
2. We consider robustness aspects of machine learning with regard to being robust to uncertainties due to lack of training data as well as unreliable sensors, and to a smaller extend also to noise found in emerging hardware.
3. We maintain a strong connection to hardware by considering resource efficiency throughout the course. While not all aspects can be called “embedded” (see LLMs), we see a strong need to keep the escalating costs of machine learning under control, no matter if we target a large-scale datacenter or a handheld device.

The main objective of this course is to help students understand these aspects, with a particular emphasis on the interplay of ML and HW.

Lecturers (current)

• Holger Fröning
• Bernhard Klein

Contents

• Review of model compression and ML processors
• CMOS scaling & non-traditional hardware technologies
• Self-attention and foundation models
• Hardware systems for LLMs
• Learned uncertainties: Bayesian deep learning based on deep ensembles and variants, Markov-Chain Monte-Carlo and Variational Inference
• HW systems for Bayesian deep learning
• Trends & Future

Requirements

Recommended is solid knowledge of PyTorch (or a similar language/tool) and the basics of computer architecture. Also, previous participation in “Embedded Machine Learning” (winter terms) is strongly recommended (but not required if equivalent knowledge is given).

Notes

• Frequency: summer term
• Among others, course qualifies for the following programs (please double-check listing within heiCO and possible specialization constraints)
  • MSc Data and Computer Science
  • MSc Computer Engineering
  • MSc Scientific Computing
  • MSc Physics

Next/current edition

• Next edition of this course is scheduled for summer 2025
• Course start: April 17, 2025, 09:00 c.t.
• Room is OMZ/INF350 basement, U014. Enter the building from the east. If you don’t see a ZITI sign when entering, you might be at the wrong entrance.
• Moodle has unrestricted enrollment. Course participation is determined by heiCO.