The 5 parts

1. The Landscape
2. The 3 Learning Paradigms
3. From Data to a Result You Trust
4. Linear Regression, Your First Model
5. Finding the Best Line

Remember Part 3 the most: Every model in this course drops into the same seven stages, and the two places students lose marks are both in there: how the data was split, and what the reported number is allowed to be.

The Tools

Tool	What it is used for	Where you meet it
Python	The language everything else is written in	Every lab, from week one
Jupyter	A notebook: code, output and prose in one file	Lab submissions
NumPy	Arrays and the arithmetic on them	Behind every model
pandas	Tables, joins, missing values	Data loading and cleaning
Matplotlib	Plots you can put in a report	Exploratory analysis
scikit-learn	The classical models, one interface	Most of this course
SciPy	Optimization, linear algebra, statistics	Under the models
seaborn	Statistical plots, in one line each	Exploratory analysis
PyTorch	Networks you define and train yourself	The deep learning weeks
TensorFlow	The same job, a different ecosystem	Named so you recognize it

Anaconda installs the whole list in one step. So make a virtual environment for each project so that one project's package versions cannot break another's.

conda create -n ds3000

The Landscape

Question

Artificial intelligence, machine learning, and deep learning get used interchangeably. What does each one actually mean, and why did the middle one take off when it did?

Goal

Place the three terms correctly inside one another, state the three ingredients every machine learning problem needs, and say what a model is really searching for.

---

Big data, and where it actually came from

Most of the time, data that is recorded is not designed to become a dataset. This is why there are missing fields, wrong units, and incorrect labels.

For example:

---

Too much to read by hand

To process all of that data (the coffee data collected in the last example) it would be inefficient to read all of the lines in the file.

So there are three solutions:

1. Look at less of it:
   • Only read a sample of a thousand rows (Your chosen sample might not contain enough fraud cases or none at all)
2. Write the rules by hand:
   • You are told what a fraud case looks like and you code it as so (if the world changes then what a fraud looks like might change, thus, rendering your code useless)
3. Have a machine find the rules:
   • Show the machine the answers you already know (the fraud and non-frauds cases) and let it figure out what separates them.
   • This course answers How?

Three words, three decades

1. Artificial intelligence
   • The effort to automate intellectual tasks normally performed by people.
   • This includes systems that have no implicit learning at all (a chess engine searching moves, or a tax program running rules an expert wrote by hand)
2. Machine learning
   • These are programs that find the rules themselves using the data given to them
   • This took off when the data and hardware improved to a point that the older mathematics could use them.
3. Deep learning
   • This is basically just MLP (Multi-layer Perceptron) in the way that it is machine learning that builds its representation in successive layers.
   • The word "deep" refers to the number of layers used, not the insight that is produced.
   • Might be called Layered or Hierarchical representation learning as the model learns that to measure as well as what to conclude.

Machine Learning took a while to take off because of the state of data and hardware at the time.

Most of the mathematics behind machine learning was published before 1920 but there wasn't enough stored data to fit a model and the hardware was not fast enough to fit it.

---

What every machine learning problem needs

1. Input data (Something measurable about each case, in a form a program can read)
2. Expected output (The right answer for each of those cases, from someone or something you trust)
3. A measure of wrong (A number saying how far the answer is form that right answer (and that can be made smaller))

For example:

Speech recognition:

• A few thousand numbers describing one slice of sound
• What a human typist heard in that slice
• How many characters the transcript got wrong

Tomorrow's power demand:

• The outside temperature that day, in degrees Celsius
• The power actually drawn that day, in megawatt hours
• How far the prediction landed from the real demand

Focus on the third one (A measure of wrong). A program that has no way to prefer one answer over another cannot get better at anything. Choosing that measure is a modeling decision which we will get to later.

PCA (Check it out later)

---

Learning is a search for better coordinates

Learning finds the parameters of a mapping that sends the inputs into a space where the classes separate.

This is the raw input space data (no mapping learning)

This is the learned mapping ("Distance form the center" was chosen by a human who could see the picture, but it is not what happens in deep learning)

❗(Picture of Learned Mapping goes here once prof makes slides interactive)

Something to note is that the only thing that changed was the coordinates of the points. This new set of coordinates is called a representation.

Working out the mapping from the data is what the learning in deep learning refers to and why those models need far more data than anything here.

Where this is already deployed

Every one of these examples has the same shape. Something measurable goes in, and one answer, that was learned from the examples, comes out.

What a dataset actually is

A dataset is basically just a table that can have three parts.

Feature: The data that you are given

Label: The outcome/result that you want

A Feature column is sometimes called: attribute, predictor, covariate, independent variable, or input

A Label column is sometimes called: target, response, outcome, dependent variable, or ground truth

A Row is sometimes called: instance, observation, example, sample, or record

The 3 Learning Paradigms

Question

What can you still learn when a label column is missing or when the only feedback you ever get is a score that arrives after the decision?

Goal

Sort a new problem into supervised, unsupervised or reinforcement learning, and say what its data would have to look like.

Supervised Learning

The data contains Feature columns AND Labeled columns

Think of the question being: "Predict something given this data"

We have 2 categories:

Regression

(See Linear Regression Code Example)

Classification

Unsupervised Learning

The data only contains Feature columns

Think of the question being: "What is already a part of this data"

We have 2 categories:

Clustering

With no label columns (the expected outputs), how do we know that the model learned correctly and produced the right groups?

Well... we don't 😭 the algorithm returns 3 groups regardless and it takes a domain expert to judge them.

Dimensionality reduction

This is NOT the same as reshaping a column (transformation).

Transformation: A log makes a skewed feature look more normal and the table keeps every column.

Reduction: deletes columns (it changes p)

Reinforcement Learning

There is no dataset at all

Here instead of a dataset, we have an Agent (the robot) and an Environment (the maze).

The Agent learns from its own experiences so each row is created by the looping of the action and the returned new state (+ reward).

NOT TALKED ABOUT IN THIS COURSE SO LOWKEY IGNORE IT!!!

From Data to a Result You Trust

Question

What actually happens between opening a data file and reporting a number you would stand behind and which of those steps is the one that quietly ruins it?

Goal

Walk a project through seven stages, split a dataset three ways for the right reason, and say why a test score stops being honest the moment you tune against it.

Should This be a Machine Learning at all?

There are 3 questions you must ask to determine whether or not a machine learning algorithm is required at all.

These 3 questions are to be answered in order.

1. Is there a pattern to find?
   • If the outcome of the problem is genuinely random, no amount of data will help.
2. Does the rule have to be discovered?
   • If the rule is not something that is already known, then write it down.
3. Do you have the answers, not just the data?
   • Get the data if you do not have it (duh) and make sure that it is labeled so it is possible to get a measure of error.

The workflow, end to end

1. Frame: This is where you want to frame the question that you want to answer.
2. Gather: This is where you either collect your data or update your dataset to include only the features, records, etc. that you need.
3. Split: Since we need to train our model, test our model and then (sometimes) validate our model, we need to split our dataset into parts and use each part for a different step (to prevent overfitting).
   • These three common sets are the Training Set, Testing Set, and Validation Set (60% / 20% / 20% split)
4. Choose: We want to use a type of model that can best fit the question (Linear Regression, Multiple Linear Regression, T-Procedures, etc.)
5. Train: Obviously this is where you give the training data to the model and allow it to learn from the data and fid the patterns within it. This happens a multitude of times to allow for better accuracy.
6. Test: (Finish Later)
7. Deploy: (Finish Later)

Linear Regression (The 1st Model)

Question

How does a computer find that pair without a person dragging a slider, and what does the search do when the surface is not a simple bowl?

Goal

Find a minimum by setting a derivative to zero, run gradient descent by hand for one step, and predict what a learning rate that is too small or too large will do.

Example: Temperature and Demand

The Simple Model:

$\hat{y} = f(x) = \theta_{0} + \theta_{1} \ x$

The predicted demand is a starting value, plus so much per degree

$x$: The feature (temperature in $\degree{C}$) $\rightarrow$ also called a covariate or a predictor $y$: The label (power demand in MWH) $\rightarrow$ also called a response, it is the answer that is already recorded $\hat{y}$: (or $f(x)$) The prediction (what the line says) $\rightarrow$ the rule that returns the estimated value. $\theta_{0}$: The y-intercept (demand at 0 $\degree{C}$) $\rightarrow$ where the line starts $\theta_{1}$: The slope (MWH per degree) $\rightarrow$ How much the prediction moves per unit of $x$.

This is the Least Squares Linear Regression:

1. Get the Means of x and y:
   • $\bar{x} =\frac{1}{n} \Sigma_{i=1}^{n}({x_i})$
   • $\bar{y} = \frac{1}{n} \Sigma_{i=1}^{n}({y_i})$
2. Get the Standard Deviations of x and y:
   • $s_{x} = \sqrt{\frac{1}{n-1}\Sigma_{i=1}^{n}(x_i - \bar{x})^{2}}$
   • $s_{y} = \sqrt{\frac{1}{n-1}\Sigma_{i=1}^{n}(y_i - \bar{y})^{2}}$
3. Get the Corr between x and y:
   • $corr(x, y) = r = \frac{1}{n-1} \Sigma_{i=1}^{n}(\frac{x_{i} - \bar{x}}{s_{x}})(\frac{y_{i} - \bar{y}}{s_{y}})$
4. Compute the Slope:
   • $m = r(\frac{s_{y}}{s_{x}})$
5. Compute the y-intercept:
   • $b = \bar{y} - x \ \bar{x}$
6. Finally, we can create our Linear prediction function:
   • $\hat{y} = mx + b$

Finding the Best Line

One issue we want to tackle is reducing the Residual Sum of Squares (RSS).

Thinking of the minimal or maximal point on a parabola, we know from calculus that the derivative at that point is equal to 0 (the slope is 0 at that point).

But the question is: Given a random point, how do we know which way to travel to reach either a local or global mimimum/maximum?

This is where Gradient Descent comes in.

Gradient Descent:

Take small steps to minimize the slope at any given point (the derivative). We want to find any local and possibly, the global minimum/maximum (when the slope is exactly 0).

To alter the step size, we want to use a hyperparameter usually called Alpha. This is the Learning Rate. WE SET THE LEARNING RATE. THE MODEL DOES NOT ALTER THIS HYPERPARAMETER!!!