Introduction

History

Artificial Intelligence (AI) was founded in 1956 at Dartmouth College. Interest in AI boomed until 1974 and funding dried up by 1980's. AI resurged around 2012 with the shift from rule-based system to data-driven learning, the evolution of deep learning algorithms (neural networks, transformers) and increased computing power.

Artificial Intelligence vs. Machine Learning vs. Deep Learning

• Deep Learning ⊄ Machine Learning ⊄ Artificial Intelligence (where ⊄ is the symbol for subset)

  • • Artificial Intelligence (AI) is the broad concept of machines learning and simulating human intelligence.

    • Machine Learning (ML) is a subset of AI where machines learn from data without explicit programming.

      • • Supervised learning: learns by using training data.

        • Unsupervised learning: learns by clustering or grouping data.

        • Reinforcement learning: Learns by interacting with the environment, for example, a robot vacuum cleaner.

    • Deep Learning (DL) is a subset of ML that uses complex neural networks to learn from vast amounts of data. It uses supervised and unsupervised learning to train deep neural networks.

Fields in Artificial Intelligence

Artificial Intelligence
|----- Behavior
|----- Robotics
|----- Cognition and Learning
|----- Fuzzy Logic
|----- Planning
|----- Knowledge representation
|----- Reason
|----- Probability
|----- Machine Learning
|----- Supervised Learning
|----- Classification
|----- SVM
|----- Decision Tree
|----- AdaBoost
|----- Naive Bayes
|----- Regression
|-----  Logistic Regression
|----- Prediction
|----- Random Forests
|----- Boosting
|----- Bagging
|----- Recommendation
|----- Collaborative Filtering
|----- Unsupervised Learning
|----- Clustering
|----- K-Means/K-Medoids
|----- Reinforcement Learning
|----- Deep Learning
|----- Transfer Learning
|----- Perception
|----- Natural Language Processing (NLP)
|----- Natural Language Understanding (NLU)
|----- Speech recognition
|----- Machine translation
|----- Text summarization
|----- Text classification
|----- Text proofreading
|----- Information extraction
|----- Natural Language Generation (NLG)
|----- Speech synthesis
|----- Computer Vision
|----- Image classification
|----- Object detection
|----- Target tracking
|----- Image segmentation

Type of AIs

• Weak AI: specializes in single tasks, for example, most ML applications including Chat GPT, semi-autonomous robots and cars (Tesla FSD), etc.

• Strong AI: can solve unseen problems, matches human intelligence, does not exist yet, for example, the fully automated robots and cars in movies.

• Super AI: super intelligent, surpass all human abilities; it does not exist yet, it's a hypothetical concept.

Neural Network

• A neural network (NN) is a type of machine learning model inspired by the human brain, consisting of interconnected nodes arranged in layers that process data to find patterns and make predictions. 

• NN has an input layer, one or more hidden layers, and an output layer, where the connections between nodes have varying strengths (weights) that are adjusted during a learning process to minimize errors and improve accuracy. 

  • • Input layer: receives raw data.

    • Hidden layers: performs computations and extract features using weights ad activation functions.

    • Output layer: Produces predictions or classifications based on learned patterns.

Deep Learning

• • The "deep" in Deep Learning refers to the depth of layers in a neural network, which consists of 3 or more hidden layers.

Concepts

• Forward propagation moves data through layers to generate an output. Backpropagation, short for Backward Propagations for Error, adjusts weight based on errors to improve accuracy. It trains neural networks by minimizing the difference between predicted and actual outputs. It works by propagating errors backward through the network, using the chain rule of calculus to compute gradients and then iteratively updating the weights and biases.

• • An activation function is a mathematical function in a neural network that determines a neuron's (node's) output based on its input. It takes the inputs, multiply with weights, add certain biases and transfers to the next layer.

  • The value produced by the activation function is the neuron's "output score." This score can vary depending on the type of activation function used:

    • • • Sigmoid: Outputs values between 0 and 1, often interpreted as probabilities in binary classification.

        • Tanh (Hyperbolic Tangent): Outputs values between -1 and 1, providing a zero-centered output.

        • ReLU (Rectified Linear Unit): Outputs 0 for negative inputs and the input value itself for positive inputs, ranging from 0 to infinity.

        • Softmax: Used in the output layer for multi-class classification, converting raw scores into a probability distribution where the sum of all outputs equals 1.

Score is the loss value during training or the confidence score of a prediction for new data. The loss score indicates the model's error (lower is better) and is a result of the loss function. Confidence scores, often derived from a final layer like a softmax function, represent the model's certainty for a specific prediction, typically in the range of 0 to 1.

• Features, also known as attributes, are the individual characteristics or pieces of information that describe the input data. These are the raw inputs fed into the neural network.

• Model parameters are defined as the internal variables of this model. They are learned or estimated purely from the data during training as every ML algorithm has mechanisms to optimize these parameters.

  • • In a simple Linear Regression model, y=ax+b, the variables a and b are the parameters of the model.

    • In a Neural Network model, the weights and biases are the parameters of the model.

    • In a Clustering model, the centroids of the clusters are the parameters of the model.

• Classes applies only to classification tasks where we want to learn a mapping function from our input features to some discrete output variables. These output variables are referred to as classes (or labels). 

• In the grad school application scenario below:

  • • features are GPAs, recommendation letters, test scores, essay, publication, etc.

    • parameters are admission requirements set in the model

    • classes are admitted (yes) or rejected (no).

Types of Neural Networks

Artificial Neural Network (ANN)

• Identify complex data patterns using interconnected layers of nodes, working with diverse inputs and adapting to unpredictable data (e.g. stock prices).

Deep Neural Network (DNN)

• A DNN is simply an ANN with multiple hidden layers—“deep” meaning more layers stacked between the input and output. A shallow NN learns simple patterns, a deep NN learns complex patterns by combining simple ones across many layers. The depth allows the network to learn complex patterns, hierarchies, and abstractions in data that shallow networks cannot.

• Applications:

  • • Computer vision

    • Self-driving cars

    • Speech recognition

    • Natural language processing

    • Fraud detection

    • Recommendation systems

    • Medical imaging

    • Game-playing agents (AlphaGo).

• Limitations:

  • • Require large amounts of data

    • Computationally expensive

    • Can be hard to interpret (“black box”)

    • Risk of overfitting if not trained carefully

• A DNN isn’t one specific model. It’s a family of deep architectures, including:

  • • CNNs (Convolutional NNs) → images

    • RNNs (Recurrent NNs) → sequences

    • Transformers → language & multimodal AI

    • Autoencoders → compression & generation

    • GANs → synthetic data generation

Recurrent Neural Network (RNN)

• Designed for sequential or time-dependent data, information where order matters.

• They have loops (connections loop back on themselves), allowing the model to remember previous inputs while processing the current one. Because of this loop, RNNs can understand: Time sequences, ordered data, context that develops over time.

• Typical neural networks process each input independently. RNNs are different — they maintain a hidden state, which is updated at every step. This lets the model keep track of what happened before.

• Applications: 

  • • Time-series forecasting

    • Speech recognition

    • Text generation

    • Language modeling

    • Machine translation

    • Music generation.

• Limitations: RNNs struggle with -

  • • Long-term dependencies

    • Vanishing gradients (the memory fades too quickly) 

    • Slow training

Convolutional Neural Network (CNN)

• Designed to process grid-like data, most commonly images.

• Extremely powerful for tasks involving visual inputs because they can automatically learn spatial patterns like edges, textures, shapes, and objects - directly from the raw pixels.

• Applications:

  • • Image classification

    • Object detection (e.g., self-driving cars)

    • Face recognition

    • Medical image analysis 

    • Video analysis

    • Image segmentation

    • Style transfer / AI art

Transformers

Transformers are a neural network architecture that learn to understand and generate human-like text by analyzing patterns in large amounts of text data. Unlike traditional models, transformers do not process data sequentially, they analyze all words at once using self-attention. They are used for performing machine learning tasks particularly in natural language processing (NLP) and computer vision.

Transformer Model Architecture

A transformer model architecture consists of two main components, which work together to capture long-term dependencies and improve translation accuracy: encoder and and decoder. The encoder extracts features from given input data, creates a numerical, context-aware representation of the input. It uses bidirectional self-attention, meaning each token can attend to all other tokens in the input sequence. The decoder uses the contextual representation produced by the encoder to generate an output sequence, typically one token at a time; it uses a masked self-attention layer, so each token can only attend to previous tokens in the output sequence, not future ones. The encoder is for understanding the input, while the decoder is for producing the output.

Attention Mechanism

Attention mechanism enables transformers to process and understand sequences without relying on recurrent or convolutional structures. It helps transformers to capture relationships between distant elements in a long sequence of data and focus on the most important parts of input data when making predictions.

Different Attention Mechanisms

Soft attention

• A differentiable attention method.

• The model assigns weights (via SoftMax) to all input positions.

• Key idea:

  • • Looks at everything, but looks at some things more than others.

• Pros:

  • • Trainable end-to-end with gradient descent.

    • Most commonly used (Transformers, seq2seq).

• Use case:

  • • Widely used in NLP tasks, text translation, summarization.

Hard attention

• Selects one part (or a small subset) of the input.

• Non-differentiable - uses sampling, not smooth weights. 

• It is trained using reinforcement learning. 

• Key idea:

  • • The model "chooses" a single location rather than blending all. 

• Training:

  • • Requires reinforcement learning. 

• Use case:

  • • Image captioning

    • Object detection

Self-attention

• Enables each input element to attend to other aspects in the same sequence.

• Attends to different position of the input sequences  

Multi-head attention

• Uses multiple attention heads to capture diverse features from different representation subspaces. 

Encoder-decoder attention

• See "Attention Architecture" paragraph below.

Hierarchical attention

• A modern type of deep neural network designed for document classification. It captures the hierarchical structure of documents by focusing on individual words and sentences using attention levels, incorporating bi-directional encoders and attentional mechanisms for classification.

Additive attention

• Uses a feed-forward neural network to calculate attention scores instead of dot products. 

BERT Model

• BERT = Bidirectional Encoder Representations from Transformers.

• BERT is a language model developed by Google that excels at understanding word relationships and context. It achieves this through pre-training on two unsupervised tasks: masked language modeling (predicting masked words) and next sentence prediction. This enables it to be fine-tuned for a wide variety of natural language processing (NLP) tasks, such as search, question answering, and text classification. 

• Unlike traditional models that process text from left to right, BERT considers the context of a word from both directions (left and right) simultaneously (Bidirectional training).

• During pre-training, BERT masks a percentage of words in a sentence and learns to predict the original words based on the surrounding context (Masked Language Model or MLM).

• BERT is also trained to predict whether a second sentence logically follows a first sentence (Next Sentence Prediction or NSP), which helps it understand sentence relationships.

• BERT is based on the Transformer architecture, which uses self-attention to weigh the importance of different words in a sentence relative to each other. 

Functional Workflow

• Input text is first preprocessed (mostly light cleaning, such as removing extra spaces or converting characters). Then it is tokenized, which means the text is broken into tokens and converted into numerical data. These numerical data are passed into an embedding layer, which turns tokens into meaningful vectors—basically, it gives words meaning in a mathematical form the model can understand. The processed numerical data then flows through the model’s neural layers to generate predictions or text outputs.