Introduction to Machine Learning and AI Architectures

Comprehensive vocabulary terms and definitions covering machine learning fundamentals, learning paradigms, and neural network architectures based on the lecture transcript.

Machine learning

A subset of artificial intelligence (AI) focused on algorithms that learn the patterns of training data and make accurate inferences about new data without explicit, hard-coded instructions.

Generalization

The fundamental goal of machine learning, referring to the translation of strong performance on training data to useful results in real-world scenarios.

AI inference

The deployment phase of an AI model where it applies patterns learned from training data to infer the correct output for a real-world task.

Deep learning

A subset of machine learning driven by large or "deep" artificial neural networks that can learn the intricate nuances of very complex data.

Artificial General Intelligence (AGI)

A type of AI often associated with science fiction, such as HAL 9000 or Ava, representing machines that can perform any intellectual task a human can.

Features

The relevant characteristics of each data point expressed numerically so they can be processed by a mathematical algorithm.

Vector embedding

A vector form representation of a data point where each element or dimension corresponds to its numerical value for a specific feature.

Feature engineering

The broader discipline of preprocessing raw data for machine learning, which includes the manual processes of feature selection and feature extraction.

Supervised learning

A learning paradigm that trains a model to predict the correct output for a given input relative to an external ground truth, such as classification or regression.

Unsupervised learning

A learning paradigm where a model discerns intrinsic patterns, dependencies, and correlations in data without an external ground truth or known correct output.

Reinforcement learning (RL)

A paradigm where an agent is trained through trial and error to evaluate its environment and take actions that will maximize a reward signal.

Ground truth

The ideal or "correct" output provided for any given input, used in supervised learning to measure and optimize model accuracy.

Loss function

A mathematical function that measures the divergence or "loss" between a model's output and the ground truth across a batch of training inputs.

Self-supervised learning

A training method where a supervisory signal is obtained directly from unlabeled data, such as autoencoders reconstructing original inputs or LLMs predicting masked words.

Clustering

A type of unsupervised learning algorithm, such as K-means or DBSCAN, that partitions unlabeled data points into groupings based on proximity or similarity.

Dimensionality reduction

Algorithms like PCA or t-SNE that reduce data complexity by representing points with a smaller number of features while preserving meaningful characteristics.

Policy

In reinforcement learning, the "thought process" or function \text{\pi} that takes a state $s$ as input and returns an action $a$, expressed as \text{\pi}(s) \rightarrow a .

Activation function

A nonlinear mathematical operation performed at each neuron or node within a neural network to enable the modeling of complex patterns.

Backpropagation

An algorithm used in deep learning to compute how each individual node contributes to the overall output of the loss function, allowing model weights to be optimized.

Universal approximators

A theoretical proof stating that for any function, there exists a neural network arrangement that can reproduce it.

Convolutional neural networks (CNNs)

Neural networks that add convolutional layers and weighted filters to extract features, primarily used for computer vision and image data.

Recurrent neural networks (RNNs)

Architecture designed for sequential data that maps a sequence of inputs to an output using a recurrent loop and an internal "memory" called the hidden state.

Transformers

An architecture for sequential data that uses an attention mechanism to selectively focus on specific parts of input data, serving as the basis for modern LLMs.

Mamba models

A neural network architecture introduced in 2023 based on a unique variation of state space models ($\text{SSMs}$), emerging as a rival to the transformer.

Machine learning operations (MLOps)

A set of practices for implementing an assembly line approach to building, deploying, and maintaining machine learning models.

Model drift

An adverse development where a deployed model undergoes changes in performance or data patterns over time, requiring monitoring and governance.