1 of 13

CSE 163

ML and Images

��Suh Young Choi

​

​

​

​

​

​

💬Before Class: Favorite board/card game?

🎵Music: Unwed Sailor

2 of 13

Machine Learning

Terms from machine learning

  • Features / labels
  • Learning algorithm
  • Model
  • Model class
  • Training set / test set
  • Parameters / Hyperparameters

2

3 of 13

ML + Images

How do we do machine learning on images?

  • Simplest: Unroll the image into a vector
  • Complex: Use other tools to extract features from the images

​

​

3

10

20

30

40

50

60

70

80

90

10

20

30

40

50

60

70

80

90

Raw Image

Unrolled Image

4 of 13

ML + Images

Pros: Simple transformation (just a call to reshape!)

Cons: It loses the idea of “neighboring” pixels (up/down)

  • Most machine learning models don’t take position of the features into account
  • This is where more complex models like convolutional neural networks come in to encode that local information as features

​

​

4

Despite these drawbacks, it can work in practice on some problems!

​

5 of 13

Neural Network

Based on how our brains �work

5

6 of 13

Example

What is the output for this neuron if the inputs are 0 for the first input and 1 for the second. The activation function is the step function (0 if negative, 1 otherwise). The bias should be subtracted from the weighted sum before applying the activation function.

6

3

-2

4

7 of 13

Example

What is the output for this neuron if the inputs are 0 for the first input and 1 for the second. The activation function is the step function (0 if negative, 1 otherwise). The bias should be subtracted from the weighted sum before applying the activation function.

7

3

-2

4

0

1

8 of 13

Example

What is the output for this neuron if the inputs are 0 for the first input and 1 for the second. The activation function is the step function (0 if negative, 1 otherwise). The bias should be subtracted from the weighted sum before applying the activation function.

8

3

-2

4

0

1

0 * 3 + 1 *(-2) - 4 = -6

9 of 13

Example

What is the output for this neuron if the inputs are 0 for the first input and 1 for the second. The activation function is the step function (0 if negative, 1 otherwise). The bias should be subtracted from the weighted sum before applying the activation function.

9

3

-2

4

0

1

0 * 3 + 1 *(-2) - 4 = -6

squish(-6) = 0

10 of 13

Image Classification

10

11 of 13

Image Classification

For a really long time, image classification was done by painstakingly crafting these features (like edge detectors), by hand.

​

This kind of worked, but we quickly hit our peak using this method.

​

Then came the buzz-word… deep learning

11

12 of 13

Image Classification

  • Is this a solved problem?
    • We get pretty decent error rates on challenges like ImageNet
  • What we can’t do
    • Sometimes can’t generalize to other real-world datasets
    • Adversarial attacks

12

13 of 13

Unsupervised

Learning

So far, we have seen supervised machine learning, where we have to explicitly shown the algorithm the labels

​

Unsupervised machine learning lets the algorithm try to learn trends on its own without providing explicit labels

​

Examples

  • Clustering
  • Outlier detection

13