Embedded Machine Learning

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 name’s “Embedded Machine Learning” stems from three different aspects:

  1. Considering that most systems to deploy ML are constrained in their resources, in particular but not limited to embedded systems, there is a strong need for resource-constrained machine learning.
  2. Furthermore, many present ML methods perform very well on data which is well curated, has no noise and is considered complete. However, when ML is deployed “in the wild” (the real world), data is scarce, possibly incomplete and noisy
  3. Last, ML’s extreme computational requirements require an understanding of the interplay among ML and HW to successfully design new ML methods. Otherwise, such methods might not prevail over time as they are not inline with current HW trends, thus they will lack their computational foundation.

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

Contents

  • Introduction/ML basics/ML processors
  • Neural networks from scratch, CONVs
  • Automatic differentiation
  • Neural architecture design (residuals, pooling, …)
  • Regularization (DropOut, L1/L2, BN, data aug.)
  • Specialized processors & safe optimizations
  • Unsafe optimizations - quantization and pruning
  • Neural architecture search
  • Time series
  • Computer vision
  • Probabilistic modeling
  • Future directions (scaling of model & data)

Requirements

Recommended is solid knowledge of C/C++/Python (either one) and the basics of computer architecture.

Notes

  • This is the inaugural edition of this course. The contents and schedule might be slightly adapted during the term.
  • Course starts 14:00 c.t on April 18 (1st week of summer term 2024). Start of exercise is to be negotiated with the teaching assistant.
  • 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.
  • This Moodle has unrestricted enrollment. Course participation is limited and will be determined later, thus enrolling in Moodle does not mean enrolling in the course for examination.

Link to Heico