AI Glossary

Machines performing tasks that typically require human intelligence

Algorithms that learn from data without explicit programming

ML using multi-layered neural networks

Training with labeled data (input + correct output)

Finding patterns in unlabeled data

Learning through trial and error with rewards/penalties

Categorizing data into predefined classes

Predicting continuous numerical values

Grouping similar data points together

Brain-inspired computing model with interconnected nodes

Model learns training data too well, performs poorly on new data

Model is too simple to capture patterns in data

Process of teaching an ML model using data

Evaluating model performance during training

Final evaluation on completely unseen data

Individual measurable property used for predictions

The correct answer in supervised learning

Probabilistic approach updating beliefs with evidence

Methods to identify cause-effect relationships

Automated model selection and tuning

Searching for the best hyperparameters

Ensemble method building models sequentially to reduce errors

Efficient, scalable gradient boosting implementation

High-performance gradient boosting by Microsoft

Gradient boosting with categorical features handling

Measure of how likely events are to occur

Collecting, analyzing, interpreting data

Probability of observing data at least as extreme under null hypothesis

Range that likely contains the true parameter value

Sequence of events where next state depends only on current state

Statistical model with hidden states and observations

Data collected sequentially over time

Classical statistical model for forecasting

Open-source forecasting tool by Meta

Centralized storage for ML features

Creating informative input variables from raw data

Collecting and importing data for processing

Measurement of data accuracy, completeness, and reliability

Centralized schema management for data streams

Tracking origin and transformations of data

Inventory of data assets with metadata

Agreements on data shape and quality between producers and consumers

Distributed streaming platform

Processing data as it arrives in real-time

Processing large chunks of data periodically

Centralized store for models and metadata

Making models available for inference

Rolling out changes to a small subset of users first

Maintain two production environments and switch traffic

Run new model in parallel without affecting users

Model performance degradation over time as data changes

Distributional changes in input data over time

Underlying relationship between features and labels changes

Observing model health and performance in production

Ability to infer system state from logs, metrics, and traces

Recording events for debugging and audit

Tracking requests across service boundaries

Service-level indicators, objectives, and agreements

Ability to reproduce results reliably

Tracking model artifacts and metadata over time

Saving model state periodically during training

Training small model (student) from large model (teacher)

Removing redundant weights or neurons to shrink models

Reducing numerical precision to shrink models and speed inference

Using lower precision where safe to speed training/inference

Open Neural Network Exchange format

NVIDIA runtime for high-performance inference

Running models on-device or near users

Machine learning on microcontrollers

Training across decentralized devices without centralizing data

Mathematical guarantee that individual data contributions remain private

Inputs intentionally perturbed to fool models

Model resilience to perturbations and distributional shifts

Understanding why a model made a decision

Shapley additive explanations for feature attribution

Local Interpretable Model-agnostic Explanations

Ensuring models do not unfairly discriminate

Systematic errors in data that lead to unfair models

Meeting regulatory and legal requirements for models

Policies and processes for responsible ML use

Protecting individuals’ personal data and rights

EU data protection regulation

Securing data by transforming it into unreadable form without keys

Protocol securing internet communications

Access control model assigning permissions by role

Securely storing and accessing credentials and keys

Recording actions for accountability and forensics

Estimating resources needed to meet demand

Measuring system or model performance under controlled tests

Analyzing resource use and hotspots in code or models

Number of requests processed per unit time

Time taken to complete a single request

Percentile-based latency metrics

Graphics Processing Unit used for parallel compute

Tensor Processing Unit by Google for ML workloads

Parallel computing platform and API for NVIDIA GPUs

Field-Programmable Gate Array for reconfigurable hardware acceleration

Application-Specific Integrated Circuit

Delay when initializing services or models not recently used

Starting from previous state to accelerate startup

Run code without managing servers

Using mostly-zero vectors to represent data efficiently

Storage and retrieval structure for vector embeddings

Approximate nearest neighbors library

Facebook AI Similarity Search library

Graph-based ANN algorithm for vector search

Systems that suggest relevant items to users

Recommends items based on other users’ preferences

Decompose user-item matrix into latent factors

Neural networks operating on graph-structured data

Graph storing entities and relationships for semantic queries

Resource Description Framework for representing graph data

Query language for RDF data

Artificially generated data for training or testing

Annotating data with ground-truth labels

Software for labeling datasets efficiently

Selectively labeling the most informative samples

Learning from small labeled and large unlabeled datasets

Learning useful representations by creating proxy tasks from raw data

Representation learning by contrasting positive and negative pairs

Learning features or embeddings that capture data structure

Size of vector representation for entities or tokens

Converts text to tokens the model understands

Subword segmentation algorithm

Tokenizer algorithm used by BERT

Set of tokens the model recognizes

Controls randomness in language model sampling

Sampling from the k most probable tokens

Sampling from smallest set of tokens whose cumulative probability ≥ p

Heuristic search keeping top sequences during decoding

Selecting highest-probability token at each step

Learning a reward function from human feedback

Reinforcement Learning from Human Feedback

Ensuring model goals match human intent

Designing AI systems to avoid harmful behavior

Adversarial testing to find model failures

Variation in time between events or packet delays

Measuring cost-efficiency of compute resources

Reducing cloud and compute spend while maintaining performance

Discounted cloud instances that can be reclaimed

Serving multiple customers on shared infrastructure

Separating workloads to prevent interference and leakage

Techniques to reduce GPU memory usage

Splitting model across devices for parallelism

Training across multiple machines or GPUs

PyTorch mechanism for synchronous multi-GPU training

Distributed training library from Uber

How often and what to save during training

Recording experiments, parameters, and metrics systematically

Platform for experiment tracking and model monitoring

Systematically removing components to measure their impact

Quick validations to ensure models behave as expected

Tests for small units of code

Tests that validate interactions between components

Full-system tests simulating real user flows

Deploy changes to small group for early validation

Reverting to a previous stable version

When a service misses its performance objective

Process for managing service disruptions

Analysis after an incident to prevent recurrence

Step-by-step instructions for operational tasks

Infrastructure for collecting and processing telemetry

Standard for observability data (logs/metrics/traces)

Monitoring system and time-series database

Visualization and dashboarding tool for metrics

Log aggregation system by Grafana Labs

Coordinated process to respond to and resolve incidents

Trained algorithm that makes predictions

Percentage of correct predictions

Of predicted positives, how many are actually positive

Of actual positives, how many did we find

Neural network specialized for processing grid-like data (images)

Neural network for sequential data with memory

Advanced RNN that remembers long-term dependencies

Attention-based architecture for sequence processing

Large language model for text generation

Language model that understands context from both directions

Two networks competing to generate realistic data

Neural network that learns compressed representations

Focusing on relevant parts of input

Non-linear function determining neuron output

Algorithm for training neural networks

Regularization technique randomly ignoring neurons

Normalizing layer inputs during training

Teaching computers to understand and work with human language

Breaking text into smaller units (tokens)

Representing words as numerical vectors

Determining emotional tone of text

Identifying and classifying named entities in text

Massive neural network trained on vast text data

LLM enhanced with external knowledge retrieval

Crafting inputs to get desired LLM outputs

Teaching computers to understand images and videos

Finding and locating objects in images

Classifying every pixel in an image

Identifying or verifying people from faces

Converting images of text to actual text

Categorizing entire images

Organized collection of data stored electronically

Language for managing relational databases

Non-relational databases for flexible data

Interface allowing software components to communicate

Web API following REST architectural principles

On-demand computing resources over the internet

Amazon's cloud computing platform

Platform for containerizing applications

Container orchestration platform

Process of moving data between systems

Database optimized for similarity search

Pre-built structure for building applications

Collection of reusable code/functions

Version control system for tracking code changes

Automating code integration and deployment

Interactive environment for code and documentation

Simplest version with core features

Demonstration that an idea is feasible

Ability to handle growth

Cost of shortcuts taken during development

Harmonic mean of precision and recall

Table showing prediction outcomes

Validating on multiple data splits

Configuration setting for learning algorithm

Step size for parameter updates during training

Optimization algorithm minimizing loss

Measures how wrong predictions are

Number of samples processed before updating weights

One complete pass through training data

Reusing pre-trained model for new task

Adapting pre-trained model to new task

Creating variations of training data

Techniques to prevent overfitting

Combining multiple models for better predictions

Ensemble of decision trees

Tree-like model of decisions

Finds optimal decision boundary

Partitions data into K clusters

Dimensionality reduction technique

Balance between underfitting and overfitting

Simplest neural network unit

Activation function: max(0, x)

S-shaped activation: 1/(1+e^-x)

Converts logits to probabilities

Downsampling layer in CNNs

Sliding window operation in CNNs

CNN with skip connections

CNN for image segmentation

Generative model learning latent representations

Real-time object detection architecture

Simplified alternative to LSTM

Encoder-decoder for sequence transformation

Algorithm for learning word embeddings

Labeling words with grammatical roles

Reducing words to dictionary form

Crude chopping to word root

Term importance in documents

Relating positions in sequence to each other

Performing tasks without specific training examples

Learning from few examples

Creating new text using language models

Automatically translating between languages

Popular document-oriented NoSQL database

In-memory key-value store

Advanced open-source relational database

Cloud object storage service

Virtual servers in AWS cloud

Serverless compute service

Query language for APIs

Persistent two-way communication protocol

Architecture of small independent services

Distributes traffic across servers

Storing frequently accessed data for fast retrieval

Distributed network serving content

Platform for hosting Git repositories

Popular code editor

Python testing framework

Google's ML framework

Facebook's ML framework

Platform for ML models and datasets

ML lifecycle management platform

Workflow orchestration platform

Iterative software development methodology

Combining development and operations

DevOps for machine learning

Comparing two versions to see which performs better

Measurable value showing progress toward goals

Ratio of profit to cost

Commitment to service availability/performance