基於內容可定址記憶體優化之樹狀集成模型加速方法
Other Title
RETENTION: Resource-Efficient Tree-Based Ensemble Model Acceleration with Content-Addressable Memory
Journal
2025臺大學士論文獎
Date Issued
2025
Author(s)
Yi-Chun Liao
Advisor
Abstract
Although deep learning has demonstrated remarkable capabilities in learning from unstructured data, modern tree-based ensemble models remain superior at extracting relevant information and learning from structured datasets. However, limited research has focused on accelerating tree-based models, whose inherent characteristics pose challenges for conventional accelerators. Content-addressable memory (CAM) offers a promising solution for accelerating tree-based models, yet it suffers from excessive memory consumption and low utilization. This work addresses these challenges by introducing RETENTION, an end-to-end framework that significantly reduces CAM capacity requirement for tree-based model inference. We propose an iterative pruning algorithm with a novel pruning criterion for bagging-based models (e.g., Random Forest) to minimize model complexity while ensuring controlled accuracy degradation. Additionally, we present a tree mapping scheme with two innovative data placement strategies to alleviate the memory redundancy associated with in-memory search. Experimental results demonstrate that RETENTION effectively reduces CAM capacity requirement, making tree-based model acceleration more feasible in resource-constrained environments.
Subjects
Tree-based machine learning
Random Forest
XGBoost, content-addressable memory
in-memory computing
pruning algorithm
Publisher
國立臺灣大學資訊工程學系
Description
獎項：傅斯年獎；指導教授：郭大維
Type
thesis
File(s)![Thumbnail Image]()
Loading...
Name
資工系 廖奕鈞(已保全).pdf
Size
1.77 MB
Format
Adobe PDF
Checksum
(MD5):8d72180e8ee959c5c30f99bd5edfa108