Intro to AI - Machine Learning

Machine Learning

Agents building models and systems based on observed data.

Artificial Intelligence

AI ≠ Machine Learning (ML); Machine Learning is a subset of AI.

Deep Learning

A part of ML involving artificial neural networks.

Learning Agent

An agent is learning if it improves performance after making observations about the world.

Model

A hypothesis about the world and software that can solve problems.

Supervised Learning

Learn a function from labeled data.

Unsupervised Learning

Learn patterns from unlabeled data.

Reinforcement Learning

Learn best actions from experience of rewards and punishments.

Labeled Data

Input-output pairs where the label is the output.

Classification

Output: finite set of values called classes or labels (e.g. true/false, sunny/rainy/cloudy).

Regression

Output: A number (e.g. temperature, which can be an integer or a real number).

Clustering

Output: Sets of similar data (based on a defined criteria).

Association Rule Mining

Output: Correlations and associations; Example: which items shoppers tend to purchase together.

Nearest Neighbors, Decision Trees, Neural Networks, Support Vector Machines, Linear Regression

Supervised learning algorithms

K Means Clustering, Hierarchical Clustering, Gaussian Mixture Models, Apriori Algorithm (association rule mining)

Unsupervised learning algorithms

Q-Learning

A reinforcement learning algorithm.

SARSA

State-Action-Reward-State-Action; a reinforcement learning algorithm.

Deep Q Network

A reinforcement learning algorithm.

Exploration

Try other options to get additional information.

Exploitation

Stay with what has given most reward.

Restaurant Waiting

Will we wait for a table in a restaurant?

Labeled Dataset

A dataset where each instance has a corresponding label.

Instances

Examples in a dataset, which can be rows in a table.

Features

Attributes or characteristics of an instance, typically represented as columns in a dataset.

Labels

The output or target variable in a supervised learning task, usually found in the last column of a dataset.

Decision Tree

A model used in machine learning for classification and regression tasks.

Attributes

The features that describe instances in a dataset.

Goal of Classification

To accurately predict the label of instances based on their features.

Rows in a Dataset

Represent individual instances in a dataset.

Columns in a Dataset

Represent features or attributes of instances in a dataset.

Last Column in Dataset

Typically contains the labels in a labeled dataset.

Example of a Labeled Dataset

A dataset containing instances with both features and corresponding labels.

Russel & Norvig

Authors of published material referenced in the slides.

Past Examples

Twelve examples with decisions made, each having ten attributes/features.

Decision

The outcome of a classification task, either 'Will Wait' or 'Will Not'.

Function (Model)

A function derived based on a dataset to predict the label of an instance with an unknown label.

Model Training

The process of training a model using features and labels from a dataset.

Model Testing

The process of testing a trained model using test features to predict labels.

Untrained Model

A model that has not yet been trained on any dataset.

Trained Model

A model that has been trained using a dataset and can make predictions.

Prediction

The output label predicted by a trained model for an instance.

Classification Models

Models used to categorize data into predefined classes, including Nearest Neighbors, Decision Trees, Random Forest, Support Vector Machines, and Neural Networks.

K in K Nearest Neighbors

The number of nearest neighbors to consider when classifying an unlabeled instance.

Instance as Datapoint

Each instance in a dataset represented as a point in a graph based on its features.

Label Representation

The label that represents most of the K nearest points in K Nearest Neighbors classification.

Feature Table

A table where each row represents a person with various features and their corresponding labels.

Patrons

Individuals represented in the feature table.

Hungry

A feature indicating whether a patron is hungry or not.

Type

A feature indicating the type of food a patron prefers.

Will Wait

A label indicating whether a patron is willing to wait for food.

Blue Label

Indicates a patron who waited for food.

Red Label

Indicates a patron who did not wait for food.

Test

Each test is based on a single feature.

Predicted Label

Eventually leads to a predicted label.

Goal of Decision Tree

A tree that most consistently leads to the correct labels (of the dataset).

Feature Selection

The feature that can best distinguish examples by their labels.

Branch via new feature

The process of splitting the decision tree based on a new feature.

Same number of 'Yes' and 'No'

Indicates a bad split in the decision tree.

Examples

Instances that are used to build the decision tree.

Dataset

A collection of examples used for training the decision tree.

Full

Indicates a complete set of data or features.

Some

Indicates a partial or incomplete set of data or features.

None

Indicates the absence of data or features.

Types of Cuisine

Examples include Thai, French, Italian, and Burger.

Decision Outcomes

The results of the tests leading to labels.

Feature Distinction

The ability of a feature to separate examples effectively.

Random Forest

An ensemble method that predicts labels based on multiple decision trees, each from a random sample of the main dataset.

Overfitting

A problem with Decision Trees where the model fits well with the training dataset but does not perform well with new instances.

Support Vector Machines (SVM)

A method where instances are treated as datapoints and features as dimensions in a hyperplane, aiming to linearly divide labeled datapoints.

Support Vectors

Points closest to the boundary in Support Vector Machines.

Artificial Neural Networks (ANN)

A model inspired by neurons and synapses in the human brain, consisting of layers of neurons connected to each other.

Input Layer

The layer in an ANN that takes in input signals, such as features.

Output Layer

The layer in an ANN that provides the output, such as labels.

Hidden Layers

Layers in an ANN that facilitate computations between the input and output layers.

Ensemble Method

A technique that combines multiple models to improve prediction accuracy.

Dimensions

Features in a dataset represented as axes in a hyperplane.

Training Dataset

The dataset used to train a model.

New Instances

Data points that the model has not seen during training.

Popular in the early 2000s

A description of the widespread use of Support Vector Machines during that time.

Layers of Neurons

The structure of an ANN where neurons are organized in layers.

Computations

The processes carried out by hidden layers in an ANN to transform inputs into outputs.

Good in Practice

A phrase describing the effectiveness of Support Vector Machines in real-world applications.

Random Sample

A subset of data selected randomly from the main dataset for training individual decision trees.

Goal of Random Forest

To predict labels based on the aggregation of predictions from multiple decision trees.

Neuron activation

A neuron is activated based on input signals, weights, thresholds, and activation function.

Back propagation

Uses back propagation to learn weights and thresholds.

Model Evaluation

To evaluate a classification model, we split our dataset into training set and test set.

Training set

used to train the model

Test set

used to evaluate the model

Accuracy

correct predictions / # total predictions

Confusion Matrix

shows correct results against predicted results for each class (i.e. possible values of label)

True Positive (TP)

Correctly predicted positive instances

False Positive (FP)

Incorrectly predicted positive instances

True Negative (TN)

Correctly predicted negative instances

False Negative (FN)

Incorrectly predicted negative instances

Number of test instances

12

Number of correct predictions

9

Accuracy example

9/12 = 0.75 or 75%

Correct predictions

are along the diagonal in the confusion matrix.

Example of Unsupervised Learning

Input: Images of animals; Output: Groups of similar images