Exploring AI

What is Artificial Intelligence?

There is no single, agreed upon answer to this question.

“The science and engineering of making intelligent machines” - John McCarthy

Machines which simulate human intelligence

What is the difference between Artificial General Intelligence (AGI) and Narrow Artificial Intelligence?

AGI can solve any problem a human can and has understanding/ability to learn comparable to human beings

NAI solves a specific problem

What is the difference between using a model in Training Mode vs Inference Mode?

Training mode is fed training instances and the output will be used to update model

Inference mode is when model is used for actual predictions

What is the difference between Training and Production?

Training process is where we are working to improve the model: uses training and interference mode

• Training when working to update the model

• Inference to validate and test

Production is when the model is fully trained and we are using it for real work: only uses inference mode

• Always in inference mode

What is the relationship between Artificial Intelligence, Machine Learning, Neural Networks, and Deep Learning?

AI —> Machine Learning —> Neural Networks —> Deep Learning

What is the difference between Symbolic AI and Subsymbolic AI?

Symbolic AI represents information as symbols or rules. → easier to understand because it explicitly represents concepts

Subsymbolic AI represents information as mathematical models

Why is Machine Learning critical to modern AI?

Almost all modern AI programs depend on some level of Machine Learning. Now we cannot define rules for complex tasks AI now carries out because the tasks we currently apply AI to are far too complex for humans to come up with a full set of explicit rules.

What is the difference between predictive, generative, and descriptive AI systems?

predictive AI forecasts future outcomes

generative AI creates original content

descriptive AI summarizes past data

Artificial General Intelligence (AGI) (Strong AI)

Term from psychology referring to doing well across a range of problems. AGI can solve any problem a human can and has understanding/ability to learn comparable to human beings.

Machine Learning

most modern AI work is actually done using machine learning. These programs learn and improve their behavior without direct modification by humans. Modern ML programs find statistical patterns in their data.

Inference

when the model is used for actual predictions

Training

• a specific mode a model (such as a neural network) operates in: when the AI fed training instances and the output will be used to update model

• serves as a reference to the entire training process: where we are working to improve the model

Production

when the model is fully trained and we are using it for real work: only uses inference mode

Predictive AI

• Designed to make predictions from data

• Learn statistical patterns from training dataset

• Apply learned patterns to previously unseen data

ex: house price, animal in a photo, weather, spam detection

Generative AI

• Generate new content (commonly text or images) based on input

• Model learns patterns from input samples provided

ex: chatbot, language translator, voice replication, music + image generation

Descriptive AI

• designed to discover previously unidentified patterns or structures in data

ex: streaming service group, system aggregates news articles

Classification

place items into pre-existing classes by sorting them into categories

Dataset

consists of individual samples that are labelled with known values. The dataset is broken into training, validation, and testing sets

Model

mathematical model designed to predict outputs from input data

Cost Function / Objective Function

compares actual output to expected output and measures how well model works

Optimizer

Updates model based on how well previous prediction worked

Narrow Artificial Intelligence (weak AI)

NAI solves a specific type of problem

Turing Test

proposed by Alan Turing in 1950. Can a computer imitate a human being well enough to fool someone.

Artificial Super Intelligence

hypothetical AI that has reasoning and understanding that exceeds humans → outperforms humans in any intellectual task: creativity and complex problem-solving ability exceed human capability

Concept Drift*

a change in the relationship between input data and what you are trying to predict changes over time, which causes a model's predictive accuracy to drop

Symbolic AI*

represents information as symbols or rules. Used for video and computer games 

Subsymbolic AI*

represents information as mathematical models

AI Winter*

The bust portion of the AI boom/bust cycle. Everyone thinks badly of AI

Data Science*

sits at the crossroads of Computer Science; Mathematics and Statistics; and Domain and Business Expertise. Data Scientists draw from all three areas to analyze business data and help business leaders make informed decisions.

Neuro-Symbolic AI*

combines the two approaches of symbolic and subsymbolic AI → neural networks can manipulate symbols

Regression*

technique used to predict and return continuous values. Analyze relationships between input variables (features) and an output variable (the target) to forecast a specific numeric outcome.

What is the difference between Supervised Learning and Unsupervised Learning?

Supervised Learning has labels with expected output

Unsupervised Learning has no labels with the expected output provided

What is Self-Supervised Learning, and why do we use it?

Self-Supervised Learning System generates its own labels using clearly defined algorithm, it is mainly used for LLMs like Chat GPT: next-word and masked-word prediction

• very useful for large datasets because it is far too large for human labelling

What is the relationship between Datasets, Samples, Features, and Labels?

Concept	Spreadsheet Analogy	Role in AI	Real-World Example (Real Estate)
Dataset	The complete file or table	The complete pool of information used to train and test the model.	The entire database of local home sales.
Sample	A single row	A single, unique instance or data point in the collection.	One specific house on Main Street.
Features	Characteristics columns	The independent input variables the model evaluates to find patterns.	Square footage, number of bedrooms, and zip code.
Label	The "Target" column	The dependent output variable or "ground truth" answer the model aims to predict.	The actual final sale price of the house.

What are the advantages of Classical Machine Learning methods over Neural Network based methods?

Classical Techniques have important advantages over Neural Networks and Deep Learning

• Models are easier to understand: less of a black box

• Less computationally expensive

• Work with less data

• Less prone to overfitting

What three mathematical specialties form the foundation for Machine Learning?

1. Statistics (and Probability)

2. Calculus

3. Linear Algebra

What is Linear Algebra?

The study of linear systems in multidimensional space (y = mx + b)

• a linear system in 2d space is a line

• a linear system in 3d space is a plane

• a linear system in n-dimensions is a hyperplane

Linear Algebraic Operations are used commonly in actual neural network operations and can also be used to modify data and come up with representations that are useful

What is a Tensor?

the basic data structure used to store our information through multi-dimensional arrays

• input data

• output results

• neural networks themselves

A good way to store many samples with different features and labels.

Suppose I’m working on a housing machine learning model. Each sample consists of the four features (square footage, yard footage, number of bedrooms, number of bathrooms) and two labels (selling price, time on market). If I wanted to represent each sample as a point in n-dimensional space (representing both inputs and outputs in the same space), what does that mean, and what would be the value of n?

n-Dimensional Space provides a theoretical framework for understanding our data. The input space alone would be n = 4, and the output space alone would be n = 2. The combined space is just their concatenation, 4 + 2 = 6. Each sample has 4 input features and 2 output labels, so if you put inputs and outputs in the same space, you get n = 6.

If I say that a model maps a point in 3-dimensional space to a point in 2-dimensional space, what does that mean? How many features does each sample have and how many labels does it have?

taking an input vector with 3 numbers and producing an output vector with 2 numbers.

• Features: 3 (the input space is 3-dimensional)

• Labels: 2 (the output space is 2-dimensional)

What is a Jupyter Notebook and why is it used?

Made with Julia, Python, and R programming languages

• can be widely installed

• accessed via the web

contain different sections split between:

• Programming Code

• Code Output (text, graphs, tables...)

• Text Sections (English)

Notebooks are widely used by

• Data Scientists

• AI Researchers

• AI Educators and Students 

but are typically not used for production work

What is Google Colab and why is it used?

combines two different technologies Jupyter Notebooks and Cloud Computing

• Jupyter notebooks stored and running on Google’s cloud infrastructure

• Runs on hardware optimized for AI work (GPUs or AI Accelerators)

• Provides pre-configured Python environments setup for AI work

Supervised Learning

Samples in dataset are labeled with expected output, generally used for Predictive Models

Unsupervised Learning

no labels (expected output) provided, generally used for Descriptive Models

Self-Supervised Learning

System generates its own labels using clearly defined algorithm, generally used for Large Language Models (e.g., ChatGPT)

• Very useful for large datasets → far too large for human labeling

Dataset (corpus)

The complete pool of information used to train and test the model.

Sample (document)

A single, unique instance or data point in the collection.

Features

occasionally used to refer to both input and output features, but more commonly refers to just the inputs

Label

the actual label on a sample, not the output of the model

Prediction

what the AI thinks the value of a label is

• Output of the model 

Classical Machine Learning

the part of the Machine Learning oval that is not inside the Neural Network oval

Linear Algebra

the study of linear systems in multidimensional space (y = mx + b)

Vectors, Matrices, and Tensors

• Vector (1D/ Rank 1): An ordered array/list of numbers, often representing magnitude and direction: a line of scalars.

• Matrix (2D / Rank 2): A rectangular grid or table of numbers arranged in rows and columns. You can think of a matrix as a collection or stacking of multiple vectors.

• Tensor (3D or Higher / Rank \(N\)): A multi-dimensional array generalizing scalars, vectors, and matrices to any number of axes (or dimensions). the basic data structure used to store our information

N-Dimensional Space

provides a theoretical framework for understanding our data

• where data is turned into lists of numbers called vectors

Google Colab

combines two different technologies Jupyter Notebooks and Cloud Computing

Jupyter Notebook

Made with Julia, Python, and R programming languages

Graphics Processing Unit (GPU)

originally created for video game 3D graphics. However, the operations needed for 3D graphics also made them quite suitable for Neural Networks. We now have GPUs specifically designed for Neural Network work

AI Accelerator

May be used as an umbrella term for any processor designed to support neural networks

Neural Processing Unit (NPU)

purpose built primarily for consumer electronic AI processing

Ground Truth

the actual value of the label

• originates from a real-world measurement of an actual data point or sample, where we know both the input features and the correct label.

MNIST

a Modified version of an earlier dataset put out by NIST (National Institute of Standards and Technology)

Semi-Supervised Learning*

a catch-all category where it is any technique which combines supervised and unsupervised

Reinforcement Learning*

Programs learns by trial an error where it receives positive and negative feedback

• There is no pre-existing dataset because data is determined by model interactions with the environment

Common Crawl*

critical dataset used by large-language models

• Dataset formed by using Bots to crawl the web

• Typically used for self-supervised learning

Traditional Machine Learning / Good-Old-Fashioned AI (GOFAI)*

older symbolic or rules-based techniques, but is considered outdated for most purposes

Linear Regression (the line of best fit)*

used to predict a continuous numerical value by finding a linear relationship between an input feature and a target output

Decision Trees*

operates like an upside-down flowchart that breaks down a complex dataset into smaller subsets

• provides a solution that can be double checked by a human expert

K-Means Clustering*

an unsupervised learning machine learning algorithm that groups unlabelled data points into k distinct, non-overlapping clusters based on feature similarity

1. Choose number of clusters in advance

2. Randomly place n centroids on the graph.

3. Label instances based on how close they are to the centroids

4. Update centroid position to center of cluster. 

5. Go to step 3 and repeat 3, 4, and 5 until centroids’ position stabilizes (or we’ve reached the preset number of iterations).

What are some of the advantages of neural-network-based machine learning over classical machine learning?

much more versatile in data patterns they work with

• given networks of sufficient size they learn much more complex patterns

• adept at finding patterns humans can’t see

able to handle wide range of input types

• much more adept with text, images, and sequence data

far more adept at generating complex outputs

What does the term “parameters” refer to when discussing Neural Networks—for example, what does “ChatGPT 3.5 has 175 billion parameters” mean?

Parameters are weights and biases in a function that are randomly generated to receive the correct output later in the function. The parameters can be changed to make a more accurate output over time.

What are weights and biases and how are they used to determine the output of a node in a neural network?

Each connection has a weight that it is multiplied by in an effort to reach the most accurate output. The “correct” weight will be learned from our training dataset. A bias is a value always added to the results

Using the diagram below showing values for inputs, weights, and biases on a neuron, walk through how we would calculate the output of the neuron.

Have weights, w, in the connections (edges) from the inputs, i. The weights and inputs are multiplied together. Pass the weights and inputs through the bias which you add to the overall cum. This then creates an overall function that is the pre-activation function. You then pass in the (wi)_n + b and then that output is the activation function

• Weights start off random, but through the training process, the weights change to accommodate the past learning

What is the primary purpose of Activation Functions in neural networks? What happens if we remove it and just use a network without Activation Functions?

The activation function is a non-linear function in which the sum of weights and biases are passed through. Without an activation function, no matter how many nodes and layers we have, we’re still limited to linear equations. 

With an activation function:

• a single node, output will somewhat resemble the activation function

• many nodes and layers we can approximate any function within a limited range.

What does it mean to have a Deep Neural Network?

A network with two or more hidden layers in which earlier nodes can look for simple relationships between inputs which allows and later nodes see how these early relationships can interact for more complex effects

What benefits does having a Deep Neural Network provide?

These all also have specialized architectures that further increase their power

What is the key characteristic that defines a Dense or Fully-Connected Neural Network?

A fully-connected in which every node in a layer is connected to every node in the previous and next layer

What kind of functions can a neural network simulate?

Any neural network with a single hidden layer simulates the functions of the human brain to, according to the Universal Approximation Theorem, simulate any continuous function within a given set of bounds to any desired level of accuracy.

What is the meaning and relationship between the following terms when used with Neural Networks: Deeper, Higher, Lower.

“Deeper” means closer to the output

“Higher” means closer to the output

“Lower” means closer to the input

What happens when our network has a high capacity, but we have a small dataset?

Overfitting occurs where the model memorizes the training data and noise rather than learning general rules

How do the number of features in each sample and the number of samples in a dataset affect a network's ability to model the data? What happens when we increase the number of features but not the number of samples?

More features allows you to find much more complex phenomena but the more data samples we need in order to fill out the space

• More features makes the process harder

• More samples makes the process easier

Node/Neuron:

a basic building block in a neural network that receives input, performs a simple math calculation, and passes the result to the next layer

Weight, Bias, Parameter

• Weights and biases combined are the parameters

• The only values that will change during the training process

Edge/Connection

nodes are connected to other nodes through a series of connections/edges

Neural Network

a set of computer algorithms modeled after the human brain that helps computers recognize patterns and solve complex problems

Input Layer

the first layer of a neural network

• purpose is to distribute values to nodes in next layer

• don’t have any parameters

• don’t have an activation function

• Lower on the network

Output Layer

the last layer of a neural network

• Typically includes parameters

• may have activation functions depending on meaning of output

• one output node for each expected value from our network

• Higher on the network

Hidden Layer

any layer hidden in the middle of the neural network that is not the output or input layer

• Determine how inputs are related to one another in ways important to the output

• More neurons means more possible combinations to consider

Deep Learning

the study of deep neural networks

Deep Neural Network

a network with two or more hidden layers

• multiple layers allow the network to combine information in more interesting ways

Dense Neural Network / Fully-Connected Neural Network

every node in a layer is connected to every node in the previous and next layer

Network Capacity

The capability of the network to learn complex patterns and relationships: determined by a combination of the number of nodes and the number of layers

Hyperparameter

They are settings determined by the neural network designer and set when the network is created

• If the network does badly, the developer can toss out these settings and redesign and retrain the network

• Can maximize network performance and minimize training time

Overfitting

too much capacity

• The network has enough capacity to start memorizing individual data points and stops focusing on underlying phenomena

Underfitting

too little capacity

• Network cannot accurately model phenomenon

• Insufficient capacity to identify or model patterns in the data

Unit*

a synonym for node and neuron

Perceptron*

may be used in place of neuron, node, or unit

• a historic, early mathematical model of a biological neuron (1958)

• a simple network of neural nodes

Multi-Level Perceptron (MLP)*

generally synonymous with a Dense Neural network

• A synonym for a common neural network architecture

Universal Approximation Theorem*

the potential power of neural networks to approximate continuous functions within a given range

• a neural network with a single hidden layer can simulate any continuous function within a given set of bounds to any desired level of accuracy.

• A single layer is often not the most efficient approach.

• A single layer may need more parameters than a deeper network to represent the same underlying function

Parameter Efficiency*

for the same number of parameters are more likely to identify complex patterns

• does depend on the task because not all tasks have the same hierarchical qualities

• Deeper networks have higher parameter efficiency

What do I mean when I say: “input order is an illusion for dense neural networks”?

There is no ordering between the nodes such that input one and four have the same relationship before training the data.