CSCI-GA.3033 

Special Topics: Efficient AI Computing: Algorithm and Implementation

fall 2025

 

The course will focus on recent advancements in the design of efficient neural networks, specifically on how to create and optimize AI models for improved performance, and resource efficiency. Students will explore essential techniques such as pruning, quantization, and model distillation across different model architectures like CNNs, RNNs, Transformers, and LLMs. These techniques aim to reduce computational complexity while preserving accuracy. Additionally, the course will cover efficient training and inference methods, including distributed computing, parallelism, and low-precision computation, essential for deploying AI on resource-constrained platforms. Lastly, students will gain a foundational understanding of computer architectures and learn how to deploy AI algorithms on actual edge devices.

 

Lecture Time: Wednesday 7:10-9:10pm EST (Zoom)

Lecture Location: Room 275, NYU Global Center for Academic and Spiritual Life (GCASL)

Readings: Course slides and papers

Suggested readings: Goodfellow, Ian. "Deep learning." (2016). https://www.deeplearningbook.org/

Evaluation Breakdown:

• Assignments (30%): total three of them, each counts 10%

• In-course quiz (10%) 

• Midterm (30%) 

• Final project (30%)

  • Proposal (1 page) 5%

  • Final presentation 15%

  • Final report 10%

Late Submission Policy: 

• A 10% deduction will be applied to the original grade if submitted within 24 hours; otherwise, a 30% deduction will be applied.

Syllabus 

Optional Paper Reading List

Course Project Instruction

Midterm Coverage

Final Project and Final Report instruction

Contact:

• For questions related to course materials, use efficientaiaccelerator@gmail.com 

  • When you send your email, please start the email title with three categories to indicate the type of the question: [Algorithm], [System], [Others].

• For other inquiries, personal matters, or emergencies, you can email me at sai.zhang@nyu.edu

Course Schedule