3 Credit Hours

(also offered as CSC 528)

JK

Jung-Eun Kim

Asst Professor

[email protected]

In this course, students read and discuss research papers about deep neural networks with a focus on not just accuracy but also resource consideration e.g., FLOPs, parameter counts, time, memory, etc. With that interest, papers about techniques to design an efficient neural network architecture, such as structured/unstructured pruning, knowledge distillation, and quantization, will be read. On top of that, other dimensional metrics of machine learning, such as trustworthiness/robustness, fairness, or privacy, will also be explored. As a seminar course, this course is dedicated to paper reading, presentation, and discussion. As this is an advanced deep learning course, students are expected to have prior implementation experiences in deep neural network architectures. Students will conduct one term project and take no exam. Students are expected to have implementation experiences on (deep) neural networks, read/present/discuss ideas from research papers, and conduct a term project and submit a term paper.

• Deep learning model architectures
• Resource metrics
• Model/Data pruning
• Model quantization
• Knowledge distilation
• Parameter-efficient fine-tuning
• Out of distribution generalization
• Memorization
• Model robustness
• Disparity in model performance

Prerequisites

Implementation experiences with (deep) neural networks.

Course Objectives

This course, as an advanced AI course, lets students read from classic to recent research papers. Students are supposed to participate into the discussions and share opinions and to ponder about the topics. Through the course, students will be able to answer the following questions:

• What techniques can make neural network architectures efficient in terms of size, inference time, parameters count, FLOPs, etc.
• What metrics/values other than accuracy can measure trustworthiness of learning models?

Course Learning Outcomes

• Students will be able to analyze and evaluate quantifiable resource consumption of neural network architectures.
• Students will be able to critique metrics to assess a network other than accuracy.
• Students will be able to synthesize various kinds of network components to build a new architecture according to their own objectives.
• Students will be able to apply the recent advancement of deep learning techniques to their real-world projects.

Course Requirements

Textbooks

None.

Created: 10/22/2026.