Electronic Computing Systems for Artificial Intelligence

Language

English
> French-friendly

Prerequisites

The complete training in electronic circuits is mandatory : ELEC1530, LELEC2531 and LELEC2532 (or equivalent) .
On top of that, it can be good to have followed advanced electronic courses( for instance LELEC2650, LELEC2570) and/or machine-learning courses (for instance LELEC2870, LELEC2885, LINFO2262).    

Main themes

Artificial intelligence is gradually becoming ubiquitous in our daily lives. As such, designing hardware for AI tasks, known as AI accelerators, and using this hardware in the context of embedded systems is becoming a major topic in electronic systems. This is for instance related to edge AI and TinyML systems, able to compute AI with limited resources.   
However, the mainstream use of AI poses several significant challenges for electronic circuits and systems: (1) how to use existing hardware (microcontrollers, AI accelerators) to compute AI tasks efficiently? (2) what are the trends and challenges in custom digital and mixed-signal hardware dedicated to AI? And (3) what are the societal and environmental impacts of AI computation?  
In this course, you will study these three challenges, at the interface between AI models, processors and system-on-chips (SoCs). After an in-depth introduction to deep learning and hands-on efficient deep neural network strategies, you will study modern AI accelerator implementations both in digital and mixed signal. Finally, you will gain perspective on the larger impacts caused by AI computing, societally and environmentally.  
This course concludes with the ELEC formation in electronic circuits and systems. 

Learning outcomes

At the end of this learning unit, the student is able to :

a. Contribution to the program’s learning outcomes:   AA1 Basic knowledge : concepts of electronic circuits (AA1.1) simulation software and CAD (AA1.2)  (architecture of digital and mixed signal AI acceleration systems, non-idealities in analog and mixed signal circuits, calibration of analog circuits, mixed-signal systems)  AA2 Engineering skills : analysis and modélisation of electronic systems and hardware/software integration   (Use of various programming languages, training and inference with AI models, compilation of AI models on a given hardware)  AA3 Research skills : gather information about existing research solutions in the topic of the project (AA3.1)  AA4 Project management  b. Other learning outcomes :   After follwing this course, engineering students in electronic circuits and systems will be able to :   Understand how current AI algorithms (in particular deep learning models) are executed on generic processors and dedicated accelerators, thanks to various model and hardware optimization techniques (in digital or mixed-signal)   Critically compare different topologies of analog and digital circuits in the context of AI acceleration, regarding task accuracy, speed, power, cost, and necessary system flexibility.  Analyse and understand the propagation of non-idealities in a mixed-signal circuit chain, and how they can be calibrated.   Understand the societal implications and the environmental aspects of AI in our society.

Content

• Introducing AI models (in particular deep learning) and optimizing their hardware computation (quantization, pruning, etc.)  
  

• Digital AI accelerator topologies 
  

• In-memory computing and mixed signal AI accelerator topologies  
  

• Non-idealities of analog circuits and self-calibration techniques 
  

• Digitally-assisted calibration of analog circuits 
  

• Emerging technologies and neuromorphic circuits for IA acceleration 
  

• Societal and environmental impacts of IA  
  

Teaching methods

The course will be organized as follows:  

• Lectures on various course topics (see contents) 
  

• A project will run for the entierty of the course, divided in three parts : (1) optimize an AI task running on a resource-contrained TinyML system, (2) optimize the architecture of a dedicated AI accelerator for given AI tasks (3) design self-compensation techniques for analog AI accelerators  
  

• Perspective lectures on emerging technologies for AI, as well as enviromental and societal impacts of AI 
  

Evaluation methods

In this course, students will be evaluated on: 

• A continuous evaluation based on group reports on the project work, to be submitted during the semester 
  

• A final oral evaluation in session 
  

Online resources

https://moodle.uclouvain.be/course/view.php?id=659

Bibliography

Chapitres de différents livres de références 
Liste de publications scientifiques récentes sur les différents sujets abordés 
Chapters of reference books 
Liste of scientific publications on the different course topics  

Teaching materials

• Coursebook, liste de publications scientifiques récentes sur les sujets abordés dans le cours

Faculty or entity

> ELEC