1 Andrew Ng, Associate Professor of Computer Science Robots and Brains
2 Who wants a robot to clean your house? [Photo Credit: iRobot]
3 Stanford AI Robot [Credit: Ken Salisbury]
4 Stanford AI Robot [Credit: Ken Salisbury]
5
6 What’s missing? Control Perception The software
7 Stanford autonomous helicopter
8 Computer GPS Accelerometers Compass
9 Stanford autonomous helicopter GPS Accelerometers Compass
10
11 Mathematical specification of helicopter [Courtesy of David Shim]
12 [Courtesy of David Shim]
13 Machine learning
14 Machine learning
15 Machine learning to fly helicopter
16 What’s missing? The software Control Perception
17 “Robot, please find my coffee mug”
18 “Robot, please find my coffee mug” Mug
19 Why is computer vision hard? But the camera sees this:
20 Computer programs (features) for vision SIFT Spin image HoG Textons Shape context GIST
21 Why is speech recognition hard? What a microphone records: “Robot, please find my coffee mug.”
22 Computer programs (features) for audio ZCR Spectrogram MFCC Rolloff Flux
23 The idea: Most of perception in the brain may be one simple program.
24 Auditory cortex learns to see Auditory Cortex The “one program” hypothesis [Roe et al., 1992]
25 Somatosensory cortex learns to see The “one program” hypothesis Somatosensory Cortex [Metin & Frost, 1989]
26 Neurons in the brain
27 Neural Network (Sparse Learning) x1x1 x2x2 x3x3 Output Layer L 1 Layer L 2 Layer L 4 Layer L 3 x4x4
28 How does the brain process images? Neuron #1 of visual cortex (model) Neuron #2 of visual cortex (model) Visual cortex looks for lines/edges.
29 Comparing to Biology Visual cortexLearning algorithm
30 Comparing to Biology Auditory cortex Learning algorithm
31 Comparing to Biology Visual Learning algorithm Brain Sound
32 Computer vision results (NORB benchmark) Neural Network: accuracy Classical computer vision (SVM): accuracy
33 Correctly found mugMistake
34 Correctly found mugMistake
35 Correctly found mugMistake
36 Correctly found mugMistake
37 Hope of progress in Artificial Intelligence
39 Option 1 BLACK
40 Machine learning Option 2
41 Comparing to Biology Learning algorithm Visual cortex [PICTURE]
42 Comparing to Biology Learning algorithm Auditory cortex
43 Missed Mugs True positivesFalse positives
44 Missed Mugs True positivesFalse positives
45 Missed Mugs True positivesFalse positives Results using non-embodied vision
46 Missed Mugs True positivesFalse positives
47 Missed Mugs True positivesFalse positives Results using non-embodied vision
48 Missed Mugs True positivesFalse positives Classifications using embodied agent
49 Missed Mugs True positivesFalse positives
50 Missed Mugs True positivesFalse positives Results using non-embodied vision
51 Missed Mugs True positivesFalse positives
52 Missed Mugs True positivesFalse positives
53 Missed Mugs True positivesFalse positives
54 Missed Mugs True positivesFalse positives