CSE 163

ML and Images�

Suh Young Choi�

🎶 Listening to: Doctor Strange soundtrack

💬 Before Class: What are you thinking about for your final project?

Announcements

No class next Monday and Wednesday (8/7 and 8/9)

• Video lectures will be posted to Canvas
• Lessons still released on Ed

​

Project Proposals due Monday 8/7

• Turned in via Gradescope
• No late work accepted, no resubmissions

​

Two more resubmission periods remaining, unless…

• If 60% of the class fills out the final course eval (releasing 8/9), then you get a third!
• You may submit any one THA and any one LR during this “bonus” period (8/16 – 8/18)

​

​

2

This Time

• Machine learning with images

Last Time

• Convolutions
• Kernel operations

3

Machine Learning, revisited

Terms from machine learning

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

4

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

​

​

5

Raw Image

Unrolled Image

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

​

​

6

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

​

Neural Network

Based on how our brains �work

7

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

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

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.

10

3

-2

4

0

1

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

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.

11

3

-2

4

0

1

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

squish(-6) = 0

Image Classification

12

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

Project Proposal Details

Multiple Datasets

• Must have at least 3 datasets that are used together in some way
• At least two of the research questions should involve at least two datasets
• Must have at least one join/merge operation between the datasets.

​

Messy Data

• Data cannot be in a file format that we have worked with in class (.csv, .shp, .json)
• Data from web-scraping, API, or requires lots of preprocessing all count as messy data
• Example: using imputation for large amounts of missing data

14

Project Proposal Details

Result Validity

• Verify the validity of your results using statistical testing or some other known testing method from domain expertise
• Any test you use must be justified and explained in the context of your data and research questions
• Clearly interpret the results of the validity tests alongside preliminary results from your analysis

​

Machine Learning

• Cannot use only the DecisionTreeRegressor or DecisionTreeClassifier pipeline from class
• Must use a new model class (scikit-learn has many!) OR
• Train several Decision Trees with differing hyperparameters

​

15

Project Proposal Details

New Library

• Must use at least one library that was not introduced in class to answer at least two research questions
• The new library may be used in tandem with one of the other challenge goals
• E.g., using Pytorch to create advanced machine learning models
• Multiple new libraries do not count as separate challenge goals
• E.g., using Scipy and plotly only count as one challenge goal

​

More details and examples can be found on the course website.

16

Before Next Time

• Complete Lesson 21
• Remember not for points, but do go towards Checkpoint Tokens
• Checkpoint 6 releasing after class today
• Project Proposals due on Monday!

Next Time

• Ethics and Data Science

​

17