Recap: Machine Learning

What we saw in the previous chapter?

(c) One Fourth Labs

Repeat the last slide of the previous chapter

A jargon cloud

How do you make sense of all the jargon?

(c) One Fourth Labs

From jargons to jars

What are the six jars of Machine Lerarning

(c) One Fourth Labs

Data data everywhere

What is the fuel of Machine Learning?

(c) One Fourth Labs

Data data everywhere

How do you feed data to machines ?

(c) One Fourth Labs

We encode all data into numbers - typically high dimension

For instance, in this course you will learn to embed image and text data as large vectors

Data entries are related - eg. given a MRI scan whether there is a tumour or not

Include a table that shows two/three MRI scans in first col, shows large vectors in second column, 1/0 for last column of whether there is tumour or not

Include a table that shows two/three reviews in first col, shows large vectors in second column, 1/0 for last column for whether review is positive or negative

Title the columns as x and y

Data data everywhere

How do you feed data to machines ?

(c) One Fourth Labs

We encode all data into numbers - typically high dimension

For instance, in this course you will learn to embed image and text data as large vectors

Data entries are related - eg. given a MRI scan whether there is a tumour or not

Include a table that shows two/three MRI scans in first col, shows large vectors in second column, 1/0 for last column of whether there is tumour or not

Include a table that shows two/three reviews in first col, shows large vectors in second column, 1/0 for last column for whether review is positive or negative

Title the columns as x and y

Data data everywhere

How do you feed data to machines ?

(c) One Fourth Labs

We encode all data into numbers - typically high dimension

For instance, in this course you will learn to embed image and text data as large vectors

Data entries are related - eg. given a MRI scan whether there is a tumour or not

Include a table that shows two/three MRI scans in first col, shows large vectors in second column, 1/0 for last column of whether there is tumour or not

Include a table that shows two/three reviews in first col, shows large vectors in second column, 1/0 for last column for whether review is positive or negative

Title the columns as x and y

Data curation

Where do I get the data from?

(c) One Fourth Labs

Source data from existing datasets

- Google datasets

- Mitesh link on datasets

- data.gov.in, etc. => Assignment: Go check out this website and formulate ML problems

Collect data yourself/others => Dataturks => Assignment: create a project, upload 5 images of sign boards, and ask five friends to label

- Take pictures of Indian dishes

- Labelling of data 

Create data specific to your problem

- Also in capstone

Data data everywhere

What is the fuel of Machine Learning?

(c) One Fourth Labs

Show data jars

Tasks

What do you do with this data?

(c) One Fourth Labs

Consider the case of Amazon with product data

Multiple tasks can be done with this: 

1. From product description to structured specs

2. From specs + reviews to writing FAQs

3. From specs + reviews + FAQs to question answering

4. From specs + reviews + personal data to recommendations

Tasks

What do you do with this data?

(c) One Fourth Labs

Consider the case of Facebook photos

Multiple tasks can be done with this: 

1. From photos identify people, places, activities

2. From posts + personal data recommend posts

3. From video detect profanity, etc. 

Tasks

What do you do with this data?

(c) One Fourth Labs

Different types of tasks: 

1. Supervised

- Classification - text or no text

- Regression - fitting bounding boxes (more later) 

2. Unsupervised

- Clustering - clustering news articles by similarity

- Generation - deep art, deep poetry

Most of the realworld ML tasks (90%) are supervised. This course will exclusively focus on this class of problems. Except for easter eggs. 

In supervised ML it is about finding y given x

Tasks

What do you do with this data?

(c) One Fourth Labs

In the dataturks labelled data, define tasks that you can perform. At least 3

// Binary classification of whether there is text

// Detect text with bounding box - is accuracy easy to define here?

Tasks

What do you do with this data?

(c) One Fourth Labs

Show data, tasks jars

What is the mathematical formulation of a task?

(c) One Fourth Labs

\( x \)

\( y \)

bat

car

dog

cat

Models

\( \left[\begin{array}{lcr} 0.2, 0.1, 0.7, ......0.8 \end{array} \right]\)

\( \left[\begin{array}{lcr} 0, 0, 1,0 \end{array} \right]\)

Now show cat, then car then ship, then dog again and keep growing the matrix

\( y  = f(x) \) [true relation, unknown]

\( \hat{y}  = \hat{f}(x) \) [our approximation]

Models

What are the choices for \( \hat{f} \) ?

(c) One Fourth Labs

- Show some points sampled from this function

- Say that there is some complex relation between x

- Naively I assumed that its y=mx + c

- no matter how I adjust m and c I can't make f and \( \hat{f} \) equal

- Let's try net function...better...better better

\( y  = mx + c \) 

\( y  = ax^2 + bx + c \) 

\( y  = \sigma(wx + b) \) 

\( y  = AlexNet(x) \) 

\( y  = \hat{f}(x) \) [our approximation]

\( \left [\begin{array}{lcr} 0.2\\ 0.1\\ 0.7\\ ....\\0.8 \end{array} \right]\)

\( \left [\begin{array}{lcr} 2.2\\ 3.1\\ 0.7\\ ....\\4.8 \end{array} \right]\)

\( x \) 

\( y \) 

\( y  = ax^3 + bx^2 + cx + d \) 

\( y  = ax^4 + bx^3 + cx + d \) 

Models

Why not just use a complex model always ?

(c) One Fourth Labs

\( \left [\begin{array}{lcr} 0.2\\ 0.1\\ 0.7\\ ....\\0.8 \end{array} \right]\)

\( \left [\begin{array}{lcr} 0.4\\ 0.2\\ 1.4\\ ....\\1.6 \end{array} \right]\)

\( x \) 

\( y \) 

This will be replaced by a simple line 

We will show animation how it will be easy to fit a line but difficult to fit 100 degree polynomial

\( y  = mx + c \) [true function, simple]

\(y = ax^{100} + bx^{99} + ... + c \) [our approximation, very complex]

• Overkill
• Harder to Learn
• Need More data

Models

What are the choices for \( \hat{f} \) ?

(c) One Fourth Labs

Add model jar

Loss Function

How do we know which model is better ?

(c) One Fourth Labs

\( \left [\begin{array}{lcr} 0.2\\ 0.1\\ 0.7\\ ....\\0.8 \end{array} \right]\)

\( \left [\begin{array}{lcr} 0.4\\ 0.2\\ 1.4\\ ....\\1.6 \end{array} \right]\)

\( x \) 

\( y \) 

\( \hat{f_1}(x)  = ax^2 + bx + c \) 

\( \hat{f_2}(x)  = ax^3 + bx^2 + cx + d \) 

\( \hat{f_3}(x)  = ax^4 + bx^3 + cx + d \) 

?

Show plots for true f and f_1 f_2 f_3... From the plots it will not be clear

but from the columns it will be clear

why is it clear? because you are computing some numbers

\( y_1 \) 

\( y_2 \) 

\( y_3 \) 

Loss Function

What does a loss function look like ?

(c) One Fourth Labs

\( \left [\begin{array}{lcr} 0.2\\ 0.1\\ 0.7\\ ....\\0.8 \end{array} \right]\)

\( \left [\begin{array}{lcr} 0.4\\ 0.2\\ 1.4\\ ....\\1.6 \end{array} \right]\)

\( x \) 

\( y \) 

\( \hat{f_1}(x)  = ax^2 + bx + c \) 

\( \hat{f_2}(x)  = ax^3 + bx^2 + cx + d \) 

\( \hat{f_3}(x)  = ax^4 + bx^3 + cx + d \) 

?

Show squared error loss

compute the error for y_1, y_2, y_3

Indeed y_2 seems to be the better model

\( y_1 \) 

\( y_2 \) 

\( y_3 \) 

Loss Function

What does a loss function look like ?

(c) One Fourth Labs

Add jar for loss function and have a recap

Who will give us the parameters ?

Learning Algorithm

How do we identify parameters of the model?

(c) One Fourth Labs

Show data on LHS

A box for learning on the RHS

A complex model equation on top of the box

Loss function at the bottom of the box 

In Plain English:

Say that this is a search problems

Simplest algorithm is to use brute force on this 3-dimensional parameter space

But now imagine what happens if you have more than 3 parameters!

Learning Algorithm

How do we identify parameters of the model?

(c) One Fourth Labs

Show data on LHS

A box for learning on the RHS

A complex model equation on top of the box

Loss function at the bottom of the box 

In Plain English:

We want to find the parameters a, b, c such that when we plugin a x into the f(x) the output should be as close to the true output

Mathematically,

Optimization problem 

Minimization function

Learning Algorithm

How do we identify parameters of the model?

(c) One Fourth Labs

Show data on LHS

A box for learning on the RHS

A complex model equation on top of the box

Loss function at the bottom of the box 

Now show images of Gradient Descent, Adam, Adagrad, etc. with citations

Learning Algorithm

How do we identify parameters of the model?

(c) One Fourth Labs

Add jar for Learning Algorithm

Evaluation

How do we compute a score for our ML model?

(c) One Fourth Labs

Show a matrix for x and y

Now add a matrix y for model predictions

Now show ticks and crosses and show we can compute accuracy (show formula)

End by saying that there are other metrics such as precision, recall, etc.

Standard evaluation (example ImageNet)

Evaluation

What are some other evaluation metrics

(c) One Fourth Labs

Show a matrix for x and y

Now add a matrix y for model predictions which is a ranked list

Now show ticks and crosses for top-1, top-5

End by saying that there are other metrics such as precision, recall, etc.

Standard evaluation (example ImageNet)

Evaluation

How is this different from loss function ?

(c) One Fourth Labs

Task is whether I should press the brake or not. I just want to know how many times I did this correctly

But to train the model I might choose to use the distance form the obstruction as a metric for training the model.

​

Evaluation

Should we learn and test on the same data?

(c) One Fourth Labs

Does it make sense to have same question as homework and exam. Why not?

This can over-estimate your performance

For an unbiased evaluation, test data should be different from train data

Typically split 80:20

Evaluation

How is this different from loss function ?

(c) One Fourth Labs

Add jar for evaluation

​

Putting it all together

How does all the jargon fit into these jars?

(c) One Fourth Labs

Show six jars

Data, democratisation, devices

Why ML is very successful?

(c) One Fourth Labs

Show six jars

Rapid progress and revolution in algorithms which have been democratized

Standardized evaluation, learning, loss, models

Standardize frameworks

You focus on getting data and formulating tasks

Typical ML effort

How to distribute your work through the six jars?

(c) One Fourth Labs

Show six jars