課程資訊
 課程名稱 深度學習之數學基礎Mathematics in Deep Learning 開課學期 110-2 授課對象 理學院  應用數學科學研究所 授課教師 黃文良 課號 MATH5255 課程識別碼 221 U8940 班次 學分 3.0 全/半年 半年 必/選修 選修 上課時間 星期四7,8,9(14:20~17:20) 上課地點 天數305 備註 總人數上限：40人 Ceiba 課程網頁 http://ceiba.ntu.edu.tw/1102MATH5255_ 課程簡介影片 核心能力關聯 本課程尚未建立核心能力關連 課程大綱 為確保您我的權利,請尊重智慧財產權及不得非法影印 課程概述 I will first review some methods for learning, then go to the technique that makes learning DNNs work. Along this line, I will focus on analysis from the perspective of optimization. The goal is to understand why and how a DNN functions. The outline roughly covers: 1. History and models of neural networks. 2. Why deep? Universality theorem for shallow/deep neural networks. 3. Optimization algorithms: back-propagation, block-coordinate descent algorithms. 4. Classification problem: dimension reduction methods, support vector machines and kernel methods. Un-supervised learning and auto-encoder. 5. Generative Adversarial Network. 6. ResNet, Mobile Net, U-nets, and RNN. 7. Using DNN to solve problems: the inverse problem, the forward inference problem, and the generalization. 課程目標 Students who are interested in analysis learn some analysis tools for DNNs. 課程要求 Linear algebra and optimization (options) 預期每週課後學習時數 Office Hours 參考書目 Given in class notes. 指定閱讀 Given in class notes. 評量方式(僅供參考)
 課程進度
 週次 日期 單元主題 第2週 2/24 NcCullock-Pitts Model and Rosenblatt's perceptron algorithm 第3週 3/03 Classification and regression trees 第4週 3/10 Support vector machine and its implication 第5週 3/17 Generalization errors 第6週 3/24 Influential instances 第7週 3/31 Universality approximation theorems for shallow and Deep NNs 第8週 4/07 Learning slow down problems 第9週 4/14 GANs and half plane cutting method 第10週 4/21 W-GAN (I) and Back-propagation algorithm (II) 第11週 4/28 The gradient unstable problem (activation functions and non-uniform weight distribution) 第12週 5/05 Take home and open reference midterm 第13週 5/12 Capturing global information from non-local mean to transformer and Bert 第14週 5/19 Dimension reduction and self-supervised learning 第15週 5/26 Autoencoder, invertible DNNs, and un-rolling algorithms. 第16週 6/02 Un-rectification(meet.google.com/ztd-bydr-cvy) 第18週 Final exam. You need to submit it before June 16th to the department (5th floor)