Chapter ML - Machine Learning

Artificial Intelligence (AI)

The capability of a machine to imitate intelligent human behavior.

NOT all artificial intelligence is machine learning.

Found in Lecture ML Part 1 - Machine Learning

Machine Learning

Giving computers the ability to learn without being explicitly programmed.

All machine learning is artificial intelligence.

Found in Lecture ML Part 1 - Machine Learning

Deep Learning

Creating machine learning models using deep neural networks.

Found in Lecture ML Part 1 - Machine Learning

Supervised Learning

Teach the machine by providing examples (with labels).

Provide examples to the machine and then let the machine learn a model that represents the examples.
Then use the model to predict values for data that are unknown.

Found in Lecture ML Part 1 - Machine Learning

Unsupervised Learning

Try to recognize patterns in the data without any examples (no labels).

Provide features to the machine and then let the machine determine the patterns that exist in the data.

Found in Lecture ML Part 1 - Machine Learning

Reinforcement Learning

Perform an action then learn based on a reward or punishment.

Reinforcement Learning considers definition of states, a set of possible actions that can be taken, and a reward score.

For a given state AND the action the machine takes, it receives a positive or negative reward.

Found in Lecture ML Part 2 - Machine Learning

Exploration vs. Exploitation in Reinforcement Learning

The core idea in reinforcement learning is trying to balance between trying new things
and using what you know.

In training, a model should have a high exploration factor so it can learn.

When deployed, a model should focus on applying the knowledge it learned.

Found in Lecture ML Part 2 - Machine Learning

Exploration factor

Parameter that determines whether you should try something new.

Found in Lecture ML Part 2 - Machine Learning

Exploitation factor

Parameter that determines whether you should use the knowledge you already have.

Found in Lecture ML Part 2 - Machine Learning

Features

The values that describe the data (e.g., x values).

Found in Lecture ML Part 1 - Machine Learning

Label

The value that you are trying to determine (e.g., y-values).

Also called the target class.

Found in Lecture ML Part 1 - Machine Learning

Classification

A type of supervised learning.

Target class has discrete values (e.g. dog, cat, mouse).

Found in Lecture ML Part 1 - Machine Learning

Regression

A type of supervised learning.

Target class has continuous values.

Found in Lecture ML Part 1 - Machine Learning

K-Means Clustering

One example of how to do unsupervised learning.

Cluster the dataset into k groups.
1. Start with k points (centroids)
2. Determine which cluster a point belongs (closest centroid)
3. Move centroid to the average of all points in that cluster
4. Repeat 2&3 until convergence

Found in Lecture ML Part 1 - Machine Learning

Centroid

Center of mass ("average" of the points).

Found in Lecture ML Part 1 - Machine Learning

Neural Networks

A neural network is a network of neurons (analogous to biological neurons in our brains).

Neural Network forms the core of every Deep Learning algorithm. These networks are represented as systems of interconnected "neurons", which send messages to each other.

Found in Lecture ML Part 3 - Machine Learning

Neurons

Each neuron has the following:
1. Multiple inputs with a weight for each input
2. An activation function
3. An output

Found in Lecture ML Part 3 - Machine Learning

Value Iteration

Finding the expected eventual reward of being in each state.

Q-Learning

Starts with an unknown set of rewards and learns each value by doing.

Neuron Step-Activation "Algorithm"

1. Multiply each input by its weight
2. Sum up all the products
3. If the sum is >= the threshold value, output 1
4. Else, output 0

Training a Neural Network

1. Start with random weights (or educated guess) and produce an incorrect result
2. Determine the error relative to correct results
3. Update the weights to account for the error
4. Repeat Steps 2 & 3 until the output converges to the correct result