Date of Publication
8-2024
Document Type
Master's Thesis
Degree Name
Bachelor of Science in Computer Engineering - Ladderized
Subject Categories
Electrical and Computer Engineering
College
Gokongwei College of Engineering
Department/Unit
Electronics And Communications Engg
Thesis Advisor
Jay Robert B. Del Rosario
Defense Panel Chair
Argel A. Bandala
Defense Panel Member
Robert Kerwin C. Billones
Donabel DeVeas-Abuan
Abstract (English)
Model compression is crucial for optimizing Convolutional Neural Networks (CNNs) for deployment on resource-constrained devices. This paper introduces Quantization with Outlier-Aware Tensor-Wise Smoothing (Q-OATS), a novel technique that integrates outlier handling with traditional model compression methods. Q-OATS addresses key challenges in activation quantization by adopting a tensor-wise approach. Unlike conventional methods that rely solely on the maximum value for activation representation, Q-OATS leverages both the mean and standard deviation, along with a migration factor, to enhance activation representation and better manage outliers. This method effectively mitigates the impact of outliers while accommodating the non-linearity and complex layer interactions inherent in CNNs, resulting in more accurate quantization. Applying Q-OATS to the ResNet-50 architecture achieved a 74% reduction in model size, from 94.42 MB to 24.11 MB, with only 0.43% accuracy degradation and a competitive inference time. Validation on AlexNet further demonstrated Q-OATS’s broad applicability across various CNN architectures. These findings underscore Q-OATS’s effectiveness in enhancing model compression by addressing activation outliers and preserving high performance. Future research should focus on applying Q-OATS to additional network architectures, diverse datasets, and tasks such as object detection, and optimizing it for deployment on edge devices.
Abstract Format
html
Abstract (Filipino)
None
Abstract Format
html
Language
English
Format
Electronic
Keywords
Neural networks (Computer science)
Recommended Citation
Sim, M. T. (2024). Q-OATS: Quantization with outlier-aware tensor-wise smoothing for CNNs. Retrieved from https://animorepository.dlsu.edu.ph/etdm_ece/37
Upload Full Text
wf_yes
Embargo Period
8-2027
