Introduction of this course 李宏毅 Hung-yi Lee

Welcome our TAs TA 信箱：ntu.mldsta@gmail.com

TAs 劉浩然 劉記良

TAs 孫凡耕 吳思霖

TAs 宋易霖 賴至得

TAs 陳力維 陳佳佑

TAs 陳冠宇 劉達融 (大助教)

What are we going to learn? 本課程內容和《機器學習》沒有重疊

and having it Deep and Structured 課程名稱解釋 機器學習 及其深層與結構化 Machine Learning and having it Deep and Structured Method

用 deep learning “硬train一發” 遇到問題，用 c4 就對了！ 用 deep learning “硬train一發”

萬事皆可 train ……

Practice of Deep Learning Previous machine learning developers Carefully design your algorithm Theoretically know its performance Deep learning Try first Many results contradict our intuition Find some reasons to explain what we observed More like chemistry Or even worse …… Ali Rahimi, Test of Time Award, NIPS 2017

Practice of Deep Learning Even a simple model can be hard to train ……

Practice of Deep Learning Interesting facts ……

Ali Rahimi, Test of Time Award, NIPS 2017

Theory of Deep Learning A network structure defines a function set Is deep better than shallow? Theory 1: Expressiveness How can we optimize by gradient descent? There are local minima …… Theory 2: Optimization Why deep network does not overfit? Although it can …… Theory 3: Generalization

Attention-based Model Theory 1 Computational Graph CNN Highway RNN Recursive Network Spatial transformer Theory 2 Batch Norm Attention-based Model Seq-to-seq Theory 3 SELU Capsule

and having it Deep and Structured 課程名稱解釋 機器學習 及其深層與結構化 Machine Learning and having it Deep and Structured Task

Output is composed of components with dependency Structured Learning Machine learning is to find a function f Regression: output a scalar Classification: output a “class” (one-hot vector) Just name a few Take human language processing and image processing as examples 1 1 1 Class 1 Class 2 Class 3 Structured Learning/Prediction: output a sequence, a matrix, a graph, a tree …… Output is composed of components with dependency

Output Sequence Machine Translation “Machine learning and having it deep and structured” “機器學習及其深層與結構化” (sentence of language 1) (sentence of language 2) Speech Recognition 感謝大家來上課” 課程名稱只能有12 個字 (speech) (transcription) Chat-bot “How are you?” “I’m fine.” (what a user says) (response of machine)

ref: https://arxiv.org/pdf/1605.05396.pdf Output Matrix Image to Image Colorization: Ref: https://arxiv.org/pdf/1611.07004v1.pdf Text to Image 畫圖 上色 “this white and yellow flower have thin white petals and a round yellow stamen” ref: https://arxiv.org/pdf/1605.05396.pdf

Reinforcement Learning Decision Making and Control Action: “right” Action: “fire” Action: “left” A sequence of decisions

Regression, Classification Structured Learning Regression, Classification

GAN RL Theory 1 Computational Graph CNN Highway RNN Recursive Network Spatial transformer Theory 2 Batch Norm Attention-based Model Seq-to-seq Theory 3 SELU Capsule GAN Generation Adversarial Network (GAN) Value-based Approach Policy-based Approach (PPO) Actor-Critic Sequence Generation Conditional Generation Imitation Learning Unsupervised Conditional Generation Auto ML RL

參考書籍 另有助教時間由助教講授作業。 Original image: http://www.danielambrosi.com/Grand-Format-Collection/i-jbhqVhS/A http://www.deeplearningbook.org/

Schedule

四次作業 HW1：深度學習流言終結者 1-1: Deep is better than shallow? 1-2: Is local minima an issue? 1-3: Is deep learning generalizable? HW2：Seq-to-seq model 2-1: Video caption generation 2-2: Chat-bot (option)

四次作業 HW3: Generative Adversarial Network (GAN) 3-1: Generation 3-2: Conditional Generation 3-3: Unsupervised Conditional Generation (option) HW4: Reinforcement learning 4-1: Policy gradient 4-2: Q-learning 4-3: Actor-critic

https://ceiba. ntu. edu. tw/modules/index. php https://ceiba.ntu.edu.tw/modules/index.php?csn=8e8d96&default_fun=syllabus

Policy 待處理: 放 cieba 幫忙組隊

成績是相對的 成績是相對的 成績是相對的 評量方式 不點名、不考試 作業一 (25%)、作業二 (25%)、作業三 (25%)、作 業四 (25%) 成績是相對的 成績是相對的 成績是相對的

組隊 每組 1 ~ 3 人，其中一人為組長 以組為單位進行所有作業 隊伍登記方法之後和作業一一起說明 組內互評：學期結束前會有組內互評，會影響成 績

上課方式 上課投影片和錄音會放到李宏毅的個人網頁上 李宏毅的個人網頁： http://speech.ee.ntu.edu.tw/~tlkagk/courses_MLSD18.h tml 社團: “Machine learning and having it deep and structured (2018 spring)” https://www.facebook.com/groups/907901649378100/ 有問題歡迎直接在 FB社團上發問 如果有同學知道答案歡迎回答 有任何和機器學習相關的想法都可以在 FB社團 上發言

提醒 深度學習必要之惡：訓練過程需要時間、調參數需要時 間 …… 作業早點開始 死線前爆氣沒有什麼幫助 健全的心靈 試著調適等待過程的焦慮 運算資源 會有 MS Azure 的贊助，但是自備運算資源更好 請勿有作弊行為 (例如：抄襲同學的作業和程式、抄襲 前人的作業和程式等等) 請遵守各作業的規定 如有規定未盡之處，則依照一般社會常識

需要的基礎能力和知識 本課程的定位為機器學習進階課程 程式能力：能夠使用某一個深度學習框架 (e.g. Tensorflow, pyTorch) 使用 Keras 無法完成所有的作業 本學期不會教深度學習框架的使用，請自學 Tensorflow https://fgc.stpi.narl.org.tw/activity/videoDetail/4b1141305d9cd2 31015d9d07dbe1002a https://fgc.stpi.narl.org.tw/activity/videoDetail/4b1141305d9cd2 31015d9d0852c5002b https://fgc.stpi.narl.org.tw/activity/videoDetail/4b1141305d9cd2 31015d9d08fb62002d pyTorch https://fgc.stpi.narl.org.tw/activity/videoDetail/4b1141305d9cd2 31015d9d0992ef0030 每次上課前會告訴大家需要先看的部分

需要的基礎能力和知識 基礎知識：希望聽課同學具備深度學習的基礎知識 《機器學習》錄影 DNN: https://www.youtube.com/watch?v=Dr-WRlEFefw Tips for DNN: https://www.youtube.com/watch?v=xki61j7z-30 CNN: https://www.youtube.com/watch?v=FrKWiRv254g RNN (Part 1): https://www.youtube.com/watch?v=xCGidAeyS4M RNN (Part 2): https://www.youtube.com/watch?v=xCGidAeyS4M Why Deep: https://www.youtube.com/watch?v=XsC9byQkUH8 Auto-encoder: https://www.youtube.com/watch?v=Tk5B4seA-AU Deep generative model (Part 1): https://www.youtube.com/watch?v=YNUek8ioAJk Deep generative model (Part 2): https://www.youtube.com/watch?v=8zomhgKrsmQ Reinforcement Learning: https://www.youtube.com/watch?v=W8XF3ME8G2I

加簽 等一下助教會公告作業 0 作業零其實是《機器學習》這門課的作業 變相擋修《機器學習》 可以用任何機器學習方法完成，不限深度學習，只要 達到要求的正確率就行 助教不會改作業零的程式 以個人為單位完成 本課程預期修課人數為 40 ~ 80 人 ，上限為 120 人 如果完成作業零人數超過上限，則按照完成作業的時 間排序 完成作業 0 後，授權碼取得方式另行公告 不會檢查程式碼