DCTrix-Net: A Low-complexity Facial Recognition Neural Net with Structured Weight Matrices

Document Type

Presentation

Location

COAS: Math Conference Room

Start Date

18-4-2025 11:15 AM

End Date

18-4-2025 11:40 AM

Description

Image recognition and classification have rapidly evolved with the advancement of machine learning algorithms. However, CNN-based learning encounters considerable challenges, including high computational demands, dependence on large volumes of labeled data, and significant storage requirements, all of which impede its overall efficiency and scalability. This presentation introduces DCTrix-Net, a novel, lightweight neural network designed specifically for efficient and accurate image recognition tasks. DCTrix-Net significantly reduces the number of trainable parameters and overall computational overhead by leveraging structured transformations based on discrete cosine transform (DCT) matrices. Experimental results show that the proposed DCTrix-Net has the lowest FLOPs and fastest inference times compared with the CNNs, RCNNs, faster RCNNs, and DCT-Net.

This is a joint work with Adam Kuzminski, Kaitlyn Cavanaugh, Hansaka Aluvihare, Krystian Stefanov, Xianqi Li, and Sirani M. Perera.

Share

COinS
 
Apr 18th, 11:15 AM Apr 18th, 11:40 AM

DCTrix-Net: A Low-complexity Facial Recognition Neural Net with Structured Weight Matrices

COAS: Math Conference Room

Image recognition and classification have rapidly evolved with the advancement of machine learning algorithms. However, CNN-based learning encounters considerable challenges, including high computational demands, dependence on large volumes of labeled data, and significant storage requirements, all of which impede its overall efficiency and scalability. This presentation introduces DCTrix-Net, a novel, lightweight neural network designed specifically for efficient and accurate image recognition tasks. DCTrix-Net significantly reduces the number of trainable parameters and overall computational overhead by leveraging structured transformations based on discrete cosine transform (DCT) matrices. Experimental results show that the proposed DCTrix-Net has the lowest FLOPs and fastest inference times compared with the CNNs, RCNNs, faster RCNNs, and DCT-Net.

This is a joint work with Adam Kuzminski, Kaitlyn Cavanaugh, Hansaka Aluvihare, Krystian Stefanov, Xianqi Li, and Sirani M. Perera.