1 of 42

OLD VERSION, USE THIS:

https://docs.google.com/presentation/d/1xcwsNSkrNtosLGbHaaIJITK8Y_AMxuxNdMS43ieTE48/edit#slide=id.p

2 of 42

What is Machine Learning?

3 of 42

Outline

Why machine learning?
What is machine learning?

How does it work?
How can we train it?
What can it do?

Going to the real world
Challenges
Wrap-up

4 of 42

Is machine learning the same as AI?

Artificial

Intelligence

Not to scale

Machine

Learning

Machine

Learning

5 of 42

Why is machine learning so exciting?

Find patterns in data that are too complicated for a human to detect.

Find solutions to problems that are difficult to solve with traditional programming.

6 of 42

Why is machine learning gaining popularity?

Solves hard problems
More accessible all the time

More data
More compute resources
Better tools

7 of 42

What is machine learning?

Computers learn to solve problems without being explicitly programmed.

Known

Input

Known Output

Model

First, training:

Input

Learned

Model

Output

Then, evaluation:

Input

Coded

Logic

Output

Classic computer science

Input

Learned

Model

Output

Machine learning

8 of 42

Given temperature, can we predict rain or snow?

?

9 of 42

How would you solve the problem?

10 of 42

Computer needs to figure it out from data.

Recorded past weather data.

Temperature in °F	Precipitation
27	Snow
52	Rain
65	Rain
10	Snow
71	Rain
...	...

11 of 42

Training the model

Find min and max temps.
For each value between min and max temps, calculate how accurately it separates “rain” temps from “snow” temps.
Return best predictor temp and its accuracy.

27	Snow
52	Rain
65	Rain
10	Snow
71	Rain
17	Snow
...	...

variable	value
min temp
max temp
current temp
current error
best predictor temp so far
lowest error so far

Training state

Training data

Watch training happen.

12 of 42

Trained model

Find min and max temps.
For each value between min and max temps, calculate how accurately it separates “rain” temps from “snow” temps.
Return best predictor temp and its accuracy.

27	Snow
52	Rain
65	Rain
10	Snow
71	Rain
17	Snow
...	...

variable	value
min temp	-1
max temp	91
current temp	91
current error	0.66
best predictor temp so far	34
lowest error so far	0.03

Training state

Training data

13 of 42

What did the model learn?

14 of 42

What does the model do with new data?

34°F

Temperature

39°F?

The model predicts snow or rain based on temperature!

15 of 42

Some terminology

A machine learning algorithm uses the features and labels in training data to train a model.

The model can then be used to evaluate new data to predict labels.

Term	Example
Features	Temperature
Labels	Rain/Snow
Model	Snow < 34° < Rain
Training data	Table of temperatures and precipitation type
Training	Try every temperature to see if it produces the most accurate results
Evaluation	Compare new input to learned temperature, predict rain or snow

16 of 42

Features, labels, model, prediction

Features:

Temperatures

Labels:

Rain/Snow

Prediction:

Rain!

Model:

Snow < 34° < Rain

Input:

39°

Model:

Snow < 34° < Rain

Training:

Evaluation:

17 of 42

More features? More dimensions.

Temperature

Temperature
Humidity

Temperature
Humidity
Altitude

18 of 42

Types of training

Supervised
Unsupervised
Semi-supervised
Reinforcement

19 of 42

Supervised Learning

Model learns based on knowing the correct output for given input.

Weather data
Mail labeled as spam
Labeled photos of animals

20 of 42

Unsupervised Learning

Model learns based only on the input.

Purchase data
Movie preference
Crime reports

21 of 42

Other training techniques

Semi-Supervised Learning

Reinforcement Learning

22 of 42

What can these models do?

Classification
Regression
Clustering

23 of 42

Classification

Predict which category input belongs in.

Example uses:

Rain or snow prediction
Spam detection
Object detection in images

24 of 42

Regression

Predict continuous, numeric output based on input.

Example uses:

Predict property value based on sales of similar homes
Project future spending based on past spending

25 of 42

Clustering

Locate naturally occurring groups within the input data.

Example uses

Security anomaly detection
Movie recommendations
Clique analysis in social networks

!?

26 of 42

How do we go from dots to the real world?

How stellar is the weather today, am I right?!

Is this message Happy or Sad?

Cat

Dog

Ewok

27 of 42

Images as numeric input

28 of 42

Words as numeric input

29 of 42

Words as numeric input

Puppy (2,4)

Kitten (4,4)

Cat (3,1)

Dog (1,1)

Puppy - Dog + Cat = Kitten

(1,3)

30 of 42

Neural networks

31 of 42

Hidden layers combine inputs

27°

60%

30m

...

Snow

Temperature

Humidity

Altitude

Input

Hidden Layer

Output

32 of 42

Each layer solves part of the problem

33 of 42

Deep Learning networks have lots of layers

34 of 42

Challenges

Sparse data
Overfitting
Biased data
Unintended proxy variables

35 of 42

What can go wrong: Sparse data

Insufficient training data for model to learn.

°F	Precipitation
47	Rain
52	Rain
65	Rain
38	Rain
...	...

36 of 42

What can go wrong: Overfitting

Model matches the noise instead of the signal in the training data.

37 of 42

What can go wrong: Biased data

Incomplete or non-representative training data results in skewed predictions.

Complex weather model trained on data in Minneapolis and applied to weather data in Albuquerque.

In one of the earliest attempted uses of supervised learning for photo identification, a detection system designed for the military was able to correctly distinguish pictures of tanks hiding among trees from pictures that had trees but no tanks in them. Although the system was 100 percent accurate in the lab, it failed in the field. Subsequent analysis revealed that the groups of pictures used for training were taken on different days with different weather and the system had merely learned to distinguish pictures based on the color of the sky.

Even if you could collect perfectly accurate data, that data will reflect societal biases. Consider the representation of different genders in different occupations. Even if your model accurately predicts the most likely career for a 6th grader based on historical data, is that the right thing to do? Left to themselves, models based on historical data tend to amplify existing biases.

38 of 42

What can go wrong: Unintended proxy variables

Keying off an unintended signal via correlated data.

We’ve talked about representing images and words as numbers. How can a human be represented by numbers? Algorithms that evaluate the probability that a human will succeed or fail at something end up relying on proxy information. For example, an individual’s score on the SAT is often used as a proxy for how that individual will succeed in college. We all know that’s not the full story, but it is still used. College admission committees know this too, which is why essays and teacher recommendations are also part of college applications. Having a human in the loop doesn’t completely fix the problem, but it helps.

Part of the benefit of having machines make decisions is that they can do much much faster than humans. In those situations, putting a human in the loop is not always feasible.

In some domains, it is explicitly illegal to discriminate individuals based on their inclusion in a protected class. Protected classes are defined by attributes like race, ethnicity, religion, gender, age, physical and mental ableness, sexual orientation, veteran status, etc. The trouble is that even when such information is intentionally left out of the input to machine learning algorithms, another piece of data may serve as a proxy for the excluded information. Classic example is the correlation of home postal code with socioeconomic status, which in this country at least, often correlates with race.

39 of 42

Current challenges

Labeled data is limited.

ML models are highly specialized to domain of training data.

Opaque ML models make it hard to determine if predictions are biased.

Experimental results are often not reproducible.

40 of 42

Wrap-up

Industry roles
Machine learning vs Artificial intelligence?

41 of 42

Industry roles

Data scientist

Data engineer

Software engineer

QA engineer for models

Data annotator

Domain expert

Computational linguist

AI/ML researcher

Ethicist

Lawyer

Marketer

Social media strategist

Public relations specialist

Product manager

Designer

HCI/UX expert

...

1 of 42

2 of 42

3 of 42

4 of 42

5 of 42

6 of 42

7 of 42

8 of 42

9 of 42

10 of 42

11 of 42

12 of 42

13 of 42

14 of 42

15 of 42

16 of 42

17 of 42

18 of 42

19 of 42

20 of 42

21 of 42

22 of 42

23 of 42

24 of 42

25 of 42

26 of 42

27 of 42

28 of 42

29 of 42

30 of 42

31 of 42

32 of 42

33 of 42

34 of 42

35 of 42

36 of 42

37 of 42

38 of 42

39 of 42

40 of 42

41 of 42

42 of 42