princeton university cos 495 instructor: yingyu liang...game google deepmind's deep q-learning...

26
Introduction to Deep Learning Princeton University COS 495 Instructor: Yingyu Liang

Upload: others

Post on 22-May-2020

9 views

Category:

Documents


0 download

TRANSCRIPT

Page 1: Princeton University COS 495 Instructor: Yingyu Liang...Game Google DeepMind's Deep Q-learning playing Atari Breakout From the paper “Playing Atari with Deep Reinforcement Learning”,

Introduction to Deep LearningPrinceton University COS 495

Instructor: Yingyu Liang

Page 2: Princeton University COS 495 Instructor: Yingyu Liang...Game Google DeepMind's Deep Q-learning playing Atari Breakout From the paper “Playing Atari with Deep Reinforcement Learning”,

What is deep learning?

• Short answer: recent buzz word

Page 3: Princeton University COS 495 Instructor: Yingyu Liang...Game Google DeepMind's Deep Q-learning playing Atari Breakout From the paper “Playing Atari with Deep Reinforcement Learning”,

Industry

• Google

• Facebook

• Microsoft

• …

• Musk

• Toyota

• Drug

• Finance

Page 4: Princeton University COS 495 Instructor: Yingyu Liang...Game Google DeepMind's Deep Q-learning playing Atari Breakout From the paper “Playing Atari with Deep Reinforcement Learning”,

Industry

• Google

Page 5: Princeton University COS 495 Instructor: Yingyu Liang...Game Google DeepMind's Deep Q-learning playing Atari Breakout From the paper “Playing Atari with Deep Reinforcement Learning”,

Industry

• Facebook

Page 6: Princeton University COS 495 Instructor: Yingyu Liang...Game Google DeepMind's Deep Q-learning playing Atari Breakout From the paper “Playing Atari with Deep Reinforcement Learning”,

Industry

• Microsoft

Page 7: Princeton University COS 495 Instructor: Yingyu Liang...Game Google DeepMind's Deep Q-learning playing Atari Breakout From the paper “Playing Atari with Deep Reinforcement Learning”,

Industry

• Elon Musk

Page 8: Princeton University COS 495 Instructor: Yingyu Liang...Game Google DeepMind's Deep Q-learning playing Atari Breakout From the paper “Playing Atari with Deep Reinforcement Learning”,

Industry

• Toyota

Page 9: Princeton University COS 495 Instructor: Yingyu Liang...Game Google DeepMind's Deep Q-learning playing Atari Breakout From the paper “Playing Atari with Deep Reinforcement Learning”,

Academy

• NIPS 2015: ~4000 attendees, double the number of NIPS 2014

Page 10: Princeton University COS 495 Instructor: Yingyu Liang...Game Google DeepMind's Deep Q-learning playing Atari Breakout From the paper “Playing Atari with Deep Reinforcement Learning”,

Academy

• Science special issue

• Nature invited review

Page 11: Princeton University COS 495 Instructor: Yingyu Liang...Game Google DeepMind's Deep Q-learning playing Atari Breakout From the paper “Playing Atari with Deep Reinforcement Learning”,

What is deep learning?

• Longer answer: machine learning framework that shows impressive performance on many Artificial Intelligence tasks

Page 12: Princeton University COS 495 Instructor: Yingyu Liang...Game Google DeepMind's Deep Q-learning playing Atari Breakout From the paper “Playing Atari with Deep Reinforcement Learning”,

Image

• Image classification• 1000 classes

Slides from Kaimin He, MSRA

Human performance: ~5%

Page 13: Princeton University COS 495 Instructor: Yingyu Liang...Game Google DeepMind's Deep Q-learning playing Atari Breakout From the paper “Playing Atari with Deep Reinforcement Learning”,

Image

• Object location

Slides from Kaimin He, MSRA

Page 14: Princeton University COS 495 Instructor: Yingyu Liang...Game Google DeepMind's Deep Q-learning playing Atari Breakout From the paper “Playing Atari with Deep Reinforcement Learning”,

Image

• Image captioning

Figure from the paper “DenseCap: Fully Convolutional Localization Networks for Dense Captioning”, by Justin Johnson, Andrej Karpathy, Li Fei-Fei

Page 15: Princeton University COS 495 Instructor: Yingyu Liang...Game Google DeepMind's Deep Q-learning playing Atari Breakout From the paper “Playing Atari with Deep Reinforcement Learning”,

Text

• Question & Answer

Figures from the paper “Ask Me Anything: Dynamic Memory Networks for Natural Language Processing ”,by Ankit Kumar, Ozan Irsoy, Peter Ondruska, Mohit Iyyer, James Bradbury, Ishaan Gulrajani, Richard Socher

Page 16: Princeton University COS 495 Instructor: Yingyu Liang...Game Google DeepMind's Deep Q-learning playing Atari Breakout From the paper “Playing Atari with Deep Reinforcement Learning”,

Game

Google DeepMind's Deep Q-learning playing Atari BreakoutFrom the paper “Playing Atari with Deep Reinforcement Learning”,by Volodymyr Mnih, Koray Kavukcuoglu, David Silver, Alex Graves, Ioannis Antonoglou,Daan Wierstra, Martin Riedmiller

Page 17: Princeton University COS 495 Instructor: Yingyu Liang...Game Google DeepMind's Deep Q-learning playing Atari Breakout From the paper “Playing Atari with Deep Reinforcement Learning”,

Game

Page 18: Princeton University COS 495 Instructor: Yingyu Liang...Game Google DeepMind's Deep Q-learning playing Atari Breakout From the paper “Playing Atari with Deep Reinforcement Learning”,

The impact

• Revival of Artificial Intelligence

• Next technology revolution?

• A big thing ongoing, should not miss

Page 19: Princeton University COS 495 Instructor: Yingyu Liang...Game Google DeepMind's Deep Q-learning playing Atari Breakout From the paper “Playing Atari with Deep Reinforcement Learning”,

Questions behind the scene

• Return of artificial neural network• What’s different

• Why get great performance

• Future development• The road to general-purpose AI?

Page 20: Princeton University COS 495 Instructor: Yingyu Liang...Game Google DeepMind's Deep Q-learning playing Atari Breakout From the paper “Playing Atari with Deep Reinforcement Learning”,

Goal of the course

• Introduction

• Key concepts

• Ticket to the party

Page 21: Princeton University COS 495 Instructor: Yingyu Liang...Game Google DeepMind's Deep Q-learning playing Atari Breakout From the paper “Playing Atari with Deep Reinforcement Learning”,

Syllabus

• Part I: machine learning basics• Linear model, Perceptron, SVM

• Multi-class

• Training by gradient descent

• overfitting

• Part II: supervised deep learning (feedforward network)

• Part III: unsupervised learning

• Part IV: deep learning in the wild

Page 22: Princeton University COS 495 Instructor: Yingyu Liang...Game Google DeepMind's Deep Q-learning playing Atari Breakout From the paper “Playing Atari with Deep Reinforcement Learning”,

Syllabus

• Part I: machine learning basics

• Part II: supervised deep learning (feedforward network)• Multiple-layer and Backpropogation

• Regularization

• Convolution

• Part III: unsupervised deep learning

• Part IV: deep learning in the wild

Page 23: Princeton University COS 495 Instructor: Yingyu Liang...Game Google DeepMind's Deep Q-learning playing Atari Breakout From the paper “Playing Atari with Deep Reinforcement Learning”,

Syllabus

• Part I: machine learning basics

• Part II: supervised deep learning (feedforward network)

• Part III: unsupervised deep learning• PCA

• Boltzmann machine, Deep Boltzmann machine

• autoencoder

• Part IV: deep learning in the wild

Page 24: Princeton University COS 495 Instructor: Yingyu Liang...Game Google DeepMind's Deep Q-learning playing Atari Breakout From the paper “Playing Atari with Deep Reinforcement Learning”,

Syllabus

• Part I: machine learning basics

• Part II: supervised deep learning (feedforward network)

• Part III: unsupervised deep learning

• Part IV: deep learning in the wild• Read papers on advanced topics

• Play with the code

• Presentation

Page 25: Princeton University COS 495 Instructor: Yingyu Liang...Game Google DeepMind's Deep Q-learning playing Atari Breakout From the paper “Playing Atari with Deep Reinforcement Learning”,

Textbook and materials

• Deep Learning:

http://www.deeplearningbook.org/

• Suggested software framework: Tensorflow• in Python

• Easy to install/use

• Can try it on your laptop

• Other software frameworks: Theano, Caffe, Torch, Marvin, …

Page 26: Princeton University COS 495 Instructor: Yingyu Liang...Game Google DeepMind's Deep Q-learning playing Atari Breakout From the paper “Playing Atari with Deep Reinforcement Learning”,

Grading

• Problem Sets (5 sets): 70%

• Design Projects: 25%

• Oral Presentation: 5%