ECE 5545 - Machine Learning Hardware and Systems

(crosslisted) CS 5775

Spring. 3 credits. Student option grading.

Prerequisite: undergraduate ECE/CS degree, programming experience, introductory ML course. Enrollment limited to: Cornell Tech students. Offered in New York City at Cornell Tech. Co-meets with ECE ECE 7785.

M. Abdelfattah.

This Master’s level course will take a hardware-centric view of machine learning systems. From constrained embedded microcontrollers to large distributed multi-GPU systems, we will investigate how these platforms run machine learning algorithms. We will look at different levels of the hardware/software/algorithm stack to make modern machine learning systems possible. This includes understanding different hardware acceleration paradigms, common hardware optimizations such as low-precision arithmetic and sparsity, compilation methodologies, model compression methods such as pruning and distillation, and multi-device federated and distributed training. Through hands-on assignments and an open-ended project, students will develop a holistic view of what it takes to train and deploy a deep neural network.

Outcome 1: Understand how machine learning algorithms run on computer systems. This includes both the hardware and the software that maps computations onto the computer chips.

Outcome 2: Apply key optimization techniques such as pruning, quantization and distillation to machine learning algorithms to improve their efficiency on different hardware platforms.

Outcome 3: Analyze the performance and efficiency of different hardware platforms with and without optimizations, and understand the impact of efficiency optimizations on the accuracy of a machine learning algorithm.

Outcome 4: Design both the hardware and software components of a machine learning computer system.