Full Stack Deep Learning Lecture 1: Intro Pieter Abbeel, Sergey - - PowerPoint PPT Presentation

Full Stack Deep Learning

Lecture 1: Intro

Pieter Abbeel, Sergey Karayev, Josh Tobin

Full Stack Deep Learning (March 2019) Pieter Abbeel, Sergey Karayev, Josh Tobin L1: Intro

Organizer Intros

1

Full Stack Deep Learning (March 2019) Pieter Abbeel, Sergey Karayev, Josh Tobin L1: Intro

Guest Instructors

Saturday Sunday Sunday Saturday

2

Full Stack Deep Learning (March 2019) Pieter Abbeel, Sergey Karayev, Josh Tobin L1: Intro

Thanks to our sponsors!

3

Full Stack Deep Learning (March 2019) Pieter Abbeel, Sergey Karayev, Josh Tobin L1: Intro

Why deep learning?

4

Full Stack Deep Learning (March 2019) Pieter Abbeel, Sergey Karayev, Josh Tobin L1: Intro

Why is AI hard to build? E.g. vision?

“C “Coffe ffee Mug” Pi Pixel el Inten ensi sity Pi Pixel el inten ensi sity y is s a a ver very y po poor repr epresen esentat ation.

[slide from Adam Coates]

5

• State-of-the-art object detection until 2012:
• Deep Supervised Learning (Krizhevsky, Sutskever, Hinton 2012; also LeCun, Bengio, Ng, Darrell,

…):

• ~1.2 million training images from ImageNet [Deng, Dong, Socher, Li, Li, Fei-Fei, 2009]

Object Detection in Computer Vision

Pieter Abbeel -- Sergey Karayev -- Josh Tobin

Input Image Hand-engineered features (SIFT, HOG, DAISY, …) Support Vector Machine (SVM) “cat” “dog” “car” … Input Image 8-layer neural network with 60 million parameters to learn “cat” “dog” “car” …

Full Stack Deep Learning (March 2019) Pieter Abbeel, Sergey Karayev, Josh Tobin L1: Intro

Neural Net Learning Image Recognition

• Training / Learning:
• Cars:
• Cats:
• Dogs:

labeled data

Machine Learning

7

* Test Time:

à à Label = Cat? Dog? Car?

Full Stack Deep Learning (March 2019) Pieter Abbeel, Sergey Karayev, Josh Tobin L1: Intro

Many Layer Neural Network

cat car dog nothing

different weights à different computation

Neural Net Training: Find the weights that minimize the difference between labels and activation.

8

Full Stack Deep Learning (March 2019) Pieter Abbeel, Sergey Karayev, Josh Tobin L1: Intro

9

Full Stack Deep Learning (March 2019) Pieter Abbeel, Sergey Karayev, Josh Tobin L1: Intro

Performance

graph credit Matt Zeiler, Clarifai

10

Full Stack Deep Learning (March 2019) Pieter Abbeel, Sergey Karayev, Josh Tobin L1: Intro

Performance

graph credit Matt Zeiler, Clarifai

11

Full Stack Deep Learning (March 2019) Pieter Abbeel, Sergey Karayev, Josh Tobin L1: Intro

Performance

graph credit Matt Zeiler, Clarifai

12

Full Stack Deep Learning (March 2019) Pieter Abbeel, Sergey Karayev, Josh Tobin L1: Intro

MS COCO Image Captioning Challenge

Karpathy & Fei-Fei, 2015; Donahue et al., 2015; Xu et al, 2015; many more

13

Full Stack Deep Learning (March 2019) Pieter Abbeel, Sergey Karayev, Josh Tobin L1: Intro

Visual QA Challenge

Stanislaw Antol, Aishwarya Agrawal, Jiasen Lu, Margaret Mitchell, Dhruv Batra, C. Lawrence Zitnick, Devi Parikh

14

Full Stack Deep Learning (March 2019) Pieter Abbeel, Sergey Karayev, Josh Tobin L1: Intro

15

Full Stack Deep Learning (March 2019) Pieter Abbeel, Sergey Karayev, Josh Tobin L1: Intro

16

Full Stack Deep Learning (March 2019) Pieter Abbeel, Sergey Karayev, Josh Tobin L1: Intro

https://www.fda.gov/NewsEvents/Newsroom/PressAnnouncements/ucm604357.htm

17

Full Stack Deep Learning (March 2019) Pieter Abbeel, Sergey Karayev, Josh Tobin L1: Intro

18

Full Stack Deep Learning (March 2019) Pieter Abbeel, Sergey Karayev, Josh Tobin L1: Intro

Speech Recognition

graph credit Matt Zeiler, Clarifai 19

Full Stack Deep Learning (March 2019) Pieter Abbeel, Sergey Karayev, Josh Tobin L1: Intro

Machine Translation

Google Neural Machine Translation (in production)

20

Full Stack Deep Learning (March 2019) Pieter Abbeel, Sergey Karayev, Josh Tobin L1: Intro

History

2000s Sparse, Probabilistic, and Energy models (Hinton, Bengio, LeCun, Ng)

Rosenblatt’s Perceptron

(Olshausen, 1996) based on history by K. Cho

21

Full Stack Deep Learning (March 2019) Pieter Abbeel, Sergey Karayev, Josh Tobin L1: Intro

Why Now?

• Data
• Compute
• Some new ideas

22

Full Stack Deep Learning (March 2019) Pieter Abbeel, Sergey Karayev, Josh Tobin L1: Intro

[Source: domo.com]

23

Full Stack Deep Learning (March 2019) Pieter Abbeel, Sergey Karayev, Josh Tobin L1: Intro

Compute

[Source: OpenAI]

24

Full Stack Deep Learning (March 2019) Pieter Abbeel, Sergey Karayev, Josh Tobin L1: Intro

Why this class?

25

Full Stack Deep Learning (March 2019) Pieter Abbeel, Sergey Karayev, Josh Tobin L1: Intro

Why this class?

26

Full Stack Deep Learning (March 2019) Pieter Abbeel, Sergey Karayev, Josh Tobin L1: Intro

n Bring people together to talk about the

challenges of working on ML projects

n We want to help those who want to

work in the field of deep learning get their dream job

Why this class?

27

Full Stack Deep Learning (March 2019) Pieter Abbeel, Sergey Karayev, Josh Tobin L1: Intro

n We talked to deep learning engineers

and leaders at Airspace, Symbio, Determined.ai, W&B, Forethought.ai, DeepScale, Google, Ancestry, Lambda Labs, Bloomberg Beta, Rent the Runway, Magic Leap, Facebook, Tesla, Uber, OpenAI, and more.

How we developed content

28

Full Stack Deep Learning (March 2019) Pieter Abbeel, Sergey Karayev, Josh Tobin L1: Intro

Schedule

29

Full Stack Deep Learning (March 2019) Pieter Abbeel, Sergey Karayev, Josh Tobin L1: Intro

Logistics

Make sure you’re in our Slack workspace (email team@fullstackdeeplearning.com if you’re not)

30

Full Stack Deep Learning (March 2019) Pieter Abbeel, Sergey Karayev, Josh Tobin L1: Intro

Questions?

31