Announcements Final exam: Friday Dec 19, 8am-11am, 1 Pimentel Closed-book, closed-notes except 1 page of notes Grading Your midterm score is replaced.

Announcements

Final exam: Friday Dec 19, 8am-11am, 1 Pimentel Closed-book, closed-notes except 1 page of notes

Grading Your midterm score is replaced by final exam score if the final score is higher

Review sessions here TuTh 12.30-2; practice final posted soon HKN Survey 1.30pm today Lecture on “Consciousness” by Christof Koch today 4pm 125 Li Ka Shing

Suggested reading: Tononi, G., & Koch, C. (2014). Consciousness: Here, There but Not Everywhere. arXiv:1405.7089.

Office hours today just 3.30-4pm1

CS 188: Artificial IntelligenceComputer Vision (all too briefly)*

Instructor: Stuart Russell --- University of California, Berkeley

Outline

Perception generally Approaches to vision Edge detection Image classification by object category

3

Perception generally Stimulus (percept) S, world W

For some function g, S = g(W) (e.g., W=scene, g=rendering, S=image) Can we do perception by inverting g?

W = g-1(S) (e.g., vision = inverse graphics) No! Percepts are usually massively ambiguous…

Percepts are evidence…

Can we just apply Bayes’ rule? P(W|S) = P(S|W) P(W) P(W) is higher for students with heads attached P(S|W) is just graphics, i.e., 1 when S=g(W), 0 otherwise

The problem is computation: Sampling from P(W) until we find a world that agrees with S is hopeless! Vision researchers have identified all sorts of cues in the image that

suggest hypotheses about (parts of) the world Need a combination of bottom-up and top-down reasoning

5

Top-down example: blind spot fill-in

http://www.colorcube.com/illusions/blndspot.htm

6

Actual image Perceived image

More examples

7http://www.skidmore.edu/~hfoley/Perc4.htm

Lightness constancy: Which has lighter pixels, A or B?

Vision cues and subsystems

9

scenes behaviors

optical flow

depth map

edges

objects

disparity

features

regions

image sequence

tracking anddata association

objectrecognition

segmentation

filters

edgedetection

matching

shape from contour

shape from stereo

shape from motion

shape from shading

scene parsing

ImagesAn image I(x,y)(t) is an array of integers [0..255] (or triples for RGB)

10

Cameras ~ 1Mpixel, human eyes ~240Mpixel x 25 fps

Edge detection

Edges in image discontinuities in scene: ⇐1) depth 2) surface orientation3) reflectance (surface markings) 4) illumination (shadows, etc.)

Edge detection contd.

1) Convolve image with spatially oriented filters (possibly multi-scale)E(x,y) = f(u,v) I(x+u,y+v) du dv

2) Label above-threshold pixels with edge orientation3) Find line segments by combining edge pixels with same orientation

12

Good, bad, and ugly

13

Illusory edges

14

Object Detection

Method 1: Deformable part-based models

Deformable part-based models

HOG features (Histogram of Oriented Gradients)

Whole-object + part-based Model of

P(part-pose | whole-pose) Efficient dynamic programming

algorithm for optimal matching Automatically learns where the

parts appear and what they look like

17

State-of-the-art results (2010)

[Girschik, Felzenszwalb, McAllester]

sofa

bottle

cat

State-of-the-art results (2010)

[Girschik, Felzenszwalb, McAllester]

person

car

horse

Method 2: Deep Learning

How Many Computers to Identify a Cat?

[Le, Ng, Dean, et al, 2012]

“Google Brain”

A typical deep learning architecture

Deep learning: Convolution network structure

Nodes in layer n take input from small patch in layer n-1

Weights are copied across all nodes in a given layer translational invariance fewer parameters =>

faster learning

23

Deep learning: Auto-encoding Stacked auto-encoder for “unsupervised” training of hidden layers from huge

collections of unlabeled images Layer 0 = input image Layer 1 = “compressed” version of Layer 0, trained with Layer 0 as output layer 2 = “compressed” version of Layer 1, trained with Layer 1 as output etc.

Then add a final layer for supervised training on categories

>0?

>0?

>0?

>0?

>0? f1

f2

f3

f1

f2

f3

ImageNet Large-Scale Visual Recognition Challenge

25

1.2 million training images in 1000 categories chosen from 14 million

images in 21,841 categories 100,000 test images “Correct” if true class is in

top-5 prediction list

ILSVRC results

26

28.2%

6.7%

Results contd.

27

Next Time

Robotics Interim results from EC3 contest Where to go next to learn more Final thoughts on the future of AI