Transcript 051811_Ng-STAN_v2x - Stanford Artificial Intelligence Laboratory

Robots and Brains
Andrew Ng, Associate Professor of Computer Science
1
Who wants a robot to clean your house?
[Photo Credit: iRobot]
2
Stanford STAIR Robot
[Credit: Ken Salisbury]
3
4
What’s missing?
The software
Control
5
Perception
Stanford autonomous helicopter
6
Stanford autonomous helicopter
GPS
Accelerometers
Compass
Computer
7
8
Computer program to fly helicopter
[Courtesy of David Shim]
9
Machine learning
12
Machine learning to fly helicopter
13
What’s missing?
The software
Control
14
Perception
“Robot, please find my coffee mug”
15
“Robot, please find my coffee mug”
Mug
Mug
Mug
Mug
Mug
16
Mug
Mug
Why is computer vision hard?
But the camera sees this:
17
Computer programs (features) for vision
18
SIFT
GIST
HoG
Shape context
Textons
Spin image
Why is speech recognition hard?
What a microphone records:
“Robot, please find my coffee mug.”
19
Computer programs (features) for audio
Spectrogram
Flux
20
MFCC
ZCR
Rolloff
The idea:
Most of perception in the brain
may be one simple program
21
The “one program” hypothesis
Auditory Cortex
Auditory cortex learns to see
[Roe et al., 1992]
22
The “one program” hypothesis
Somatosensory Cortex
Somatosensory cortex learns to see
[Roe et al., 1992]
23
Neurons in the brain
24
Neural Network
x1
x2
Output
x3
Layer L4
x4
Layer L1
25
Layer L3
Layer L2
How does the brain process images?
Primary visual cortex looks for “edges.”
Neuron #1 of visual cortex
(model)
26
Neuron #2 of visual cortex
(model)
Comparing to Biology
Visual cortex
[PICTURE]
27
Learning algorithm
Comparing to Biology
Auditory cortex
[PICTURE]
28
Learning algorithm
[PICTURE]
Computer vision results (NORB benchmark)
29
Classical computer
vision (SVM):
Neural Network:
accuracy
accuracy
Missed Mugs
30
True positives
False positives
Missed Mugs
31
True positives
False positives
Missed Mugs
40
True positives
False positives
Missed Mugs
41
True positives
False positives
Hope of progress in
Artificial Intelligence
Email: [email protected]
42