國立中正大學數學系

暨應用數學碩士班、統計科學碩士班

學    術   演    講

基於新型可解釋神經元進行特徵選擇並降低模型成本之研究

A Study on Feature Selection and Computational Cost Reduction Based on Novel Explainable Neurons

邱勝敏 博士

Dr. Chiu, Sheng-Min

瑞鼎科技 Raydium Corp

Abstract 近年來，深度學習模型 (DLM) 在各個領域得到了廣泛應用。學者在使用DLMs時通常會提供豐 富的數據集，以使模型更好地學習目標任務。而普遍來說DLMs運算成本過高的問題是阻礙其 廣泛被普羅大眾使用的最大主因。本研究提出可解釋性的神經元進行特徵選擇並降低模型成本 之研究為克服此難題，通過在 DLM 中引入用於特徵選擇和成本降低的可解釋神經元來應對這 一挑戰。本研究側重於二個不同的領域：工業應用和時空數據庫，工業應用中，我們提出了輕 量化的刀具磨耗預測框架。時空數據庫中則針對房地產售價預測及人流預測的多種問題進行特 徵選擇及降低成本的研究。每個部份根據其資料特性得到專門地處理及設計，所有提出的方法 都通過實驗進行了徹底的檢查、模擬、分析和驗證。實驗結果證明了所提出的方法在解決與 DLM 相關的高計算成本方面的有效性。

Deep learning models (DLMs) have gained widespread application in various domains in recent years. Researchers typically provide rich datasets to enable DLMs to better learn target tasks. However, the high computational cost of DLMs remains a significant barrier to their widespread adoption. This thesis proposes a research study that addresses this challenge by introducing interpretable neurons for feature selection and cost reduction in DLMs. This study focuses on two distinct domains: industrial applications and spatio-temporal databases. In the domain of industrial applications, a lightweight framework for predicting tool wear is proposed. In the spatio-temporal database domain, research is conducted on various aspects, including feature selection and cost reduction for real estate price prediction and crowd flow prediction. Each of these domains is extensively discussed in dedicated chapters throughout the thesis. All proposed methods are thoroughly examined, simulated, analyzed, and validated through experiments. The experimental results demonstrate the efficacy of the proposed approaches in addressing the high computational costs associated with DLMs

日 期:113 年 12 月4日(星期三) 16:10~17:00

地 點：本校數學館 527 教室（嘉義縣民雄鄉大學路 168 號）

茶 會：15:30~16:00 數學館四樓 409 室舉行

歡迎參加 敬請公佈

交通資訊及校內地圖