Scribd Logo
Upload

Welcome to Scribd!

  • 10 views

    Introduction (v2)

    Uploaded by

    Cheng Zihan

    Copyright:

    © All Rights Reserved
    Download as PDF, TXT or read online from Scribd

    Introduction (v2)

    Uploaded by

    Cheng Zihan
    10 views43 pages

    Original Title

    introduction (v2)

    Copyright

    © © All Rights Reserved

    Available Formats

    PDF, TXT or read online from Scribd

    Did you find this document useful?

    Is this content inappropriate?

    Copyright:

    © All Rights Reserved
    Download as PDF, TXT or read online from Scribd
    Download as pdf or txt
    10 views43 pages

    Introduction (v2)

    Uploaded by

    Cheng Zihan

    Copyright:

    © All Rights Reserved
    Download as PDF, TXT or read online from Scribd
    Download as pdf or txt
    of 43

    請保持社交距離

    本課程有直播可以線上觀看

    直播連結 課程網頁 加簽表單


    https://youtu.be/eKgDxp-_A0c
    機器學習
    課程速覽

    李宏毅 Hung-yi Lee


    課程內容
    • 本課程共十五講，課程錄影已經放在課程網頁上
    • 每一講都有一個對應的作業，作業內容之後上課
    時間公布
    • 上課時間會講新的內容 (與作業無關)

    課程網頁
    什麼是機器學習？
    Machine Learning
    ≈ Looking for Function
    • Speech Recognition

    f( ) = “How are you”


    • Image Recognition
    f( ) = “Cat”
    • Playing Go

    f( ) = “5-5”
    (next move)
    Different types of Functions

    This course focuses on Deep Learning.

    Neural
    Network
    Different types of Functions
    regression
    scalar
    vector

    classification
    e.g.,
    f
    image
    Matrix

    Sequence
    e.g., speech, text text image
    HW1: COVID-19 Case Prediction
    regression
    scalar
    vector

    classification
    e.g.,
    f
    image
    Matrix

    Sequence
    e.g., speech, text text image
    HW2: Phoneme Classification
    regression
    scalar
    vector

    classification
    e.g.,
    f
    image
    Matrix

    Sequence
    e.g., speech, text text image
    HW3: Image Classification
    regression
    scalar
    vector

    classification
    e.g.,
    f
    image
    Matrix

    Sequence
    e.g., speech, text text image
    HW4: Speaker Classification
    regression
    scalar
    vector

    classification
    e.g.,
    f
    image
    Matrix

    Sequence
    e.g., speech, text text image
    HW5: Machine Translation
    regression
    scalar
    vector

    classification
    e.g.,
    f
    image
    Matrix

    Sequence
    e.g., speech, text text image
    HW6: Anime Face Generation
    regression
    scalar
    vector

    classification
    e.g.,
    f
    image
    Matrix

    Sequence
    e.g., speech, text text image
    教機器的種種方法
    Supervised Learning Lecture 1 - 5

    Pokémon
    or
    Digimon
    Training Data

    Pokémon Pokémon Digimon Digimon


    labels
    Lecture 7:
    Self-supervised Learning

    It is not efficient to collect data for each task.


    Lecture 7:
    Self-supervised Learning
    unlabeled
    images

    Are they the


    same?

    Develop general
    purpose knowledge
    Are they the
    Pre-train same?
    Lecture 7:
    Self-supervised Learning

    Fine-tune

    Pokémon Digimon

    Develop general
    purpose knowledge
    Fine-tune
    Pre-train apple orange

    Downstream Tasks
    Lecture 7:
    Self-supervised Learning
    BERT Pre-trained Model vs. Downstream Tasks
    (Foundation Model)

    Operating Systems Applications


    BERT
    340M
    parameters

    Attack on Titan
    Source of image:
    https://leemeng.tw/attack_on_bert_transfer_learning_in_nlp.html
    Spoiler Alert
    BERT
    Bertolt
    340M
    Hoover
    parameters

    Attack on Titan
    Source of image:
    https://leemeng.tw/attack_on_bert_transfer_learning_in_nlp.html
    GPT-3 Source: https://youtu.be/wJJnjzNlMws

    T5
    GPT-2

    BERT
    ELMo
    Lecture 6:
    Generative Adversarial Network

    𝒙 Function 𝒚

    𝒙𝟏 𝒚𝟐
    𝒙 𝟑 𝒙𝟕 𝒚𝟒 𝒚𝟏𝟎
    𝒙𝟓
    𝟗 𝒚𝟖 𝒚𝟔
    𝒙
    unpaired
    Unsupervised Abstractive
    Summarization
    https://arxiv.org/abs/1810.02851
    document summary

    Unsupervised Translation
    https://arxiv.org/abs/1710.04087
    https://arxiv.org/abs/1710.11041
    Language 1 Language 2

    Unsupervised ASR
    https://arxiv.org/abs/1804.00316
    https://arxiv.org/abs/1812.09323
    https://arxiv.org/abs/1904.04100
    https://arxiv.org/abs/2105.11084
    Audio Text
    Lecture 12:
    Reinforcement Learning (RL)
    Human label
    ? Pokémon

    It is challenging to label data in some tasks.

    Human label
    ? “3-3”?

    We can know the results are good or not. → RL


    進階課題
    不只是追求正確率 …
    Lecture 8: Anomaly Detection
    This is a
    “Pokémon”.

    This is a
    “Digimon”.

    I do not know
    Lecture 9: Explainable AI

    This is a
    “Pokémon”.

    Because .…
    Classifier

    Why do you think this image is a Pokémon?


    Lecture 9: Explainable AI

    Testing Accuracy: 98.4% Amazing!!!!!!


    Lecture 9: Explainable AI
    Lecture 9: Explainable AI
    Lecture 9: Explainable AI
    • All the images of Pokémon are PNG, while most
    images of Digimon are JPEG.

    loading the files

    png files have transparent transparent background


    background becomes black
    Machine discriminates Pokémon and Digimon
    based on the background colors.
    I will let you know the story
    after fixing the mistake. ☺
    Lecture 10: Model Attack
    Benign Image Attacked Image

    Tiger Cat Star Fish


    0.64 1.00
    Lecture 10: Model Attack
    =
    Benign Image Attacked Image

    50x -

    Tiger Cat Star Fish


    0.64 1.00
    Lecture 11: Domain Adaptation

    Training
    Data

    Testing
    Data
    99.5% 57.5%

    The results are from: http://proceedings.mlr.press/v37/ganin15.pdf


    Lecture 13: Network Compression
    smaller

    Deploying ML models in resource- Too Big!


    constrained environments

    Lower latency, Privacy, etc.


    Lecture 14: Life-long Learning
    I can solve I can solve I can solve
    task 1. tasks 1&2. tasks 1&2&3.

    Learning Learning Learning


    Task 1 Task 2 Task 3

    This is the target of life-long learning.


    What is the challenge?
    學習如何學習
    Meta Learning = Learn to Learn
    Lecture 15: Meta learning
    Few-shot learning is usually achieved by meta-learning.
    Learn to classify

    Learning
    Algorithm
    I hope you enjoy this course!
    台大電機系機器學習課程 YouTube 頻道

    You might also like