A practical dictionary of 512 AI, machine learning, mathematics, and programming terms from HerbDev’s AI Vocabulary app. Start with a definition, explore an example, then follow the concepts behind it.
Linked concepts open in a new tab. Audio reads the definition and example.
512 terms
Page 1 of 22 · Showing 1–24
Artificial Intelligence · Level 2 of 5
A-Star Search
A path search that combines cost so far with estimated remaining cost.
With suitable heuristic conditions, A-star can find an optimal path.
Example
Navigation ranks routes by distance traveled plus distance estimated to remain.
Listen to the definition and example
Audio transcript
A-Star Search.
A path search that combines cost so far with estimated remaining cost.
With suitable heuristic conditions, A-star can find an optimal path.
For example: Navigation ranks routes by distance traveled plus distance estimated to remain.
Explore this concept
Why it matters
These distinctions help you separate a system’s task, representation, and reasoningmethod.
Try answering before looking back at the definition.
Research Vocabulary · Level 2 of 5
Ablation Study
An experiment removing or altering components to estimate their contribution.
Interactions between components can complicate interpretation.
Example
A model is retrained without attention to assess its role.
Listen to the definition and example
Audio transcript
Ablation Study.
An experiment removing or altering components to estimate their contribution.
Interactions between components can complicate interpretation.
For example: A model is retrained without attention to assess its role.
Explore this concept
Why it matters
This helps you read papers critically and distinguish evidence from claims.
Try answering before looking back at the definition.
Model Evaluation · Level 2 of 5
Accuracy
The fraction of evaluated predictions that are correct.
It can hide poor performance on rare classes.
Example
Ninety correct labels out of one hundred give 90 percent accuracy.
Listen to the definition and example
Audio transcript
Accuracy.
The fraction of evaluated predictions that are correct.
It can hide poor performance on rare classes.
For example: Ninety correct labels out of one hundred give 90 percent accuracy.
Explore this concept
A useful analogy
The share of all answers marked correct on a test.
Why it matters
This helps you ask what a reported score counts and which errors it can hide.
Try answering before looking back at the definition.
Reinforcement Learning · Level 1 of 5
Action
A choice available to an agent that can affect its environment.
Actions can be discrete or continuous.
Example
A robot chooses a steering angle.
Listen to the definition and example
Audio transcript
Action.
A choice available to an agent that can affect its environment.
Actions can be discrete or continuous.
For example: A robot chooses a steering angle.
Explore this concept
Why it matters
This helps you separate what an agent observes, chooses, and optimizes.
The word can refer to a signal, while activation function names the rule producing it.
Example
A hidden unit outputs 0.7 for one example.
Listen to the definition and example
Audio transcript
Activation.
A value produced by a unit or layer after a transformation.
The word can refer to a signal, while activation function names the rule producing it.
For example: A hidden unit outputs 0.7 for one example.
Explore this concept
Why it matters
This is part of the vocabulary used to read network diagrams and understand parameter updates.
Try answering before looking back at the definition.
Neural Networks · Level 2 of 5
Activation Function
A transformation applied to a unit's signal, often to introduce nonlinearity.
Without nonlinear operations, stacked affine layers collapse into one affine map.
Example
A rectifier replaces negative values with zero.
Listen to the definition and example
Audio transcript
Activation Function.
A transformation applied to a unit's signal, often to introduce nonlinearity.
Without nonlinear operations, stacked affine layers collapse into one affine map.
For example: A rectifier replaces negative values with zero.
Explore this concept
Why it matters
This is part of the vocabulary used to read network diagrams and understand parameter updates.
Try answering before looking back at the definition.
Machine Learning · Level 2 of 5
Active Learning
Selecting examples for labeling to improve a model efficiently.
The learner asks for labels it expects to find informative.
Example
A system sends its most uncertain images to an annotator.
Listen to the definition and example
Audio transcript
Active Learning.
Selecting examples for labeling to improve a model efficiently.
The learner asks for labels it expects to find informative.
For example: A system sends its most uncertain images to an annotator.
Explore this concept
Why it matters
This helps you choose and evaluate a learning setup that matches the data and task.
Try answering before looking back at the definition.
Reinforcement Learning · Level 4 of 5
Actor-Critic
A reinforcement-learning design combining a policy with a value estimator.
The critic helps evaluate or reduce variance in the actor's updates.
Example
One network chooses actions while another estimates their future value.
Listen to the definition and example
Audio transcript
Actor-Critic.
A reinforcement-learning design combining a policy with a value estimator.
The critic helps evaluate or reduce variance in the actor's updates.
For example: One network chooses actions while another estimates their future value.
Explore this concept
Why it matters
This helps you separate what an agent observes, chooses, and optimizes.
Try answering before looking back at the definition.
Optimization · Level 2 of 5
Adam
An optimizer using moving estimates of gradient means and squared gradients.
It rescales updates coordinatewise and corrects initial moment-estimate bias.
Example
A trainer maintains two running gradient statistics for each parameter.
Listen to the definition and example
Audio transcript
Adam.
An optimizer using moving estimates of gradient means and squared gradients.
It rescales updates coordinatewise and corrects initial moment-estimate bias.
For example: A trainer maintains two running gradient statistics for each parameter.
Explore this concept
A useful analogy
A walker that smooths recent directions and adjusts the scale separately for each coordinate.
Why it matters
This helps you understand what parameter updates are trying to improve and when they can fail.
Technical detail
Adam tracks m_t = β₁m_(t−1)+(1−β₁)g_t and v_t = β₂v_(t−1)+(1−β₂)g_t², then uses bias-corrected moments.
Common misconception
Adam’s adaptive steps do not remove the need to choose a learning rate.
Decoupling distinguishes its shrinkage from an L2 term inside Adam's gradient.
Example
A trainer applies parameter shrinkage independently of its moment-scaled update.
Listen to the definition and example
Audio transcript
AdamW.
Adam with weight decay applied separately from adaptive gradient scaling.
Decoupling distinguishes its shrinkage from an L2 term inside Adam's gradient.
For example: A trainer applies parameter shrinkage independently of its moment-scaled update.
Explore this concept
Why it matters
This helps you understand what parameter updates are trying to improve and when they can fail.
Try answering before looking back at the definition.
Responsible AI / Safety · Level 2 of 5
Adversarial Example
An input crafted to cause a model to behave incorrectly or undesirably.
Attack assumptions and allowed changes define the threat model.
Example
A small image perturbation changes a classifier's prediction.
Listen to the definition and example
Audio transcript
Adversarial Example.
An input crafted to cause a model to behave incorrectly or undesirably.
Attack assumptions and allowed changes define the threat model.
For example: A small image perturbation changes a classifier's prediction.
Explore this concept
Why it matters
This helps you identify a concrete failure mode or evaluate the limits of a proposed control.
Try answering before looking back at the definition.
Agents & Tool Use · Level 2 of 5
Agent
A system that chooses and executes actions using observations and goals.
An LLM-based agent typically combines a model with tools and control logic.
Example
An assistant checks a calendar before proposing a meeting time.
Listen to the definition and example
Audio transcript
Agent.
A system that chooses and executes actions using observations and goals.
An LLM-based agent typically combines a model with tools and control logic.
For example: An assistant checks a calendar before proposing a meeting time.
Explore this concept
Why it matters
This helps you separate model suggestions from the software actions that execute them.
Try answering before looking back at the definition.
Agents & Tool Use · Level 2 of 5
Agent Loop
A repeated observe-decide-act cycle with a stopping condition.
Unbounded loops can waste resources or repeat mistakes.
Example
A controller alternates tool results and model decisions until completion.
Listen to the definition and example
Audio transcript
Agent Loop.
A repeated observe-decide-act cycle with a stopping condition.
Unbounded loops can waste resources or repeat mistakes.
For example: A controller alternates tool results and model decisions until completion.
Explore this concept
Why it matters
This helps you separate model suggestions from the software actions that execute them.
Try answering before looking back at the definition.
Algorithms · Level 1 of 5
Algorithm
A specified procedure for carrying out a computation or solving a problem.
An algorithm's behavior can be analyzed separately from one implementation.
Example
A search procedure repeatedly halves a sorted range.
Listen to the definition and example
Audio transcript
Algorithm.
A specified procedure for carrying out a computation or solving a problem.
An algorithm's behavior can be analyzed separately from one implementation.
For example: A search procedure repeatedly halves a sorted range.
Explore this concept
Why it matters
This helps you compare problem-solving methods and their computational tradeoffs.
Try answering before looking back at the definition.
Responsible AI / Safety · Level 2 of 5
Alignment
Making system behavior better match intended goals, constraints, or human preferences.
The intended values and evaluation criteria must be specified.
Example
A model is trained and evaluated to follow task instructions within defined boundaries.
Listen to the definition and example
Audio transcript
Alignment.
Making system behavior better match intended goals, constraints, or human preferences.
The intended values and evaluation criteria must be specified.
For example: A model is trained and evaluated to follow task instructions within defined boundaries.
Explore this concept
Why it matters
This helps you identify a concrete failure mode or evaluate the limits of a proposed control.
Try answering before looking back at the definition.
Data & Datasets · Level 2 of 5
Annotation
Adding structured labels or metadata to data.
Annotation guidelines determine what annotators are asked to distinguish.
Example
A person marks the boundaries of a tumor in an image.
Listen to the definition and example
Audio transcript
Annotation.
Adding structured labels or metadata to data.
Annotation guidelines determine what annotators are asked to distinguish.
For example: A person marks the boundaries of a tumor in an image.
Explore this concept
Why it matters
This helps you prevent misleading experiments caused by data preparation and split mistakes.
Try answering before looking back at the definition.
Programming · Level 2 of 5
API (API)
A defined way for software components to interact.
An API can be local; it does not necessarily involve the internet.
Example
An application calls an operating-system notification method.
Listen to the definition and example
Audio transcript
API.
A defined way for software components to interact.
An API can be local; it does not necessarily involve the internet.
For example: An application calls an operating-system notification method.
Explore this concept
Why it matters
This helps you read implementation code and describe its behavior precisely.
An application shares work among CPU, GPU, and specialized accelerators on one chip.
Listen to the definition and example
Audio transcript
Apple Silicon.
Apple-designed system-on-chip processors integrating multiple compute components.
Capabilities vary across chips and devices.
For example: An application shares work among CPU, GPU, and specialized accelerators on one chip.
Explore this concept
Why it matters
This helps you identify whether compute, memory, or runtime support limits execution.
Try answering before looking back at the definition.
Embeddings & Vector Search · Level 3 of 5
Approximate Nearest Neighbor
Searching for close vectors with a speed or memory tradeoff against exact results.
Approximation can miss some true nearest neighbors.
Example
An index searches a small fraction of a million vectors.
Listen to the definition and example
Audio transcript
Approximate Nearest Neighbor.
Searching for close vectors with a speed or memory tradeoff against exact results.
Approximation can miss some true nearest neighbors.
For example: An index searches a small fraction of a million vectors.
Explore this concept
Why it matters
This helps you understand why a retrieval system finds or misses related information.
Argument.
A concrete value supplied when calling a function.
Arguments fill the function's parameter positions or named bindings.
For example: The value five is passed into a squaring routine.
Explore this concept
Why it matters
This helps you distinguish a computation’s contract, inputs, outputs, and behavior.
Array.
An indexed collection of elements.
Mutability and element-type restrictions depend on the language.
For example: A list of scores is accessed by position.
Explore this concept
Why it matters
This helps you read implementation code and describe its behavior precisely.
Try answering before looking back at the definition.
Artificial Intelligence · Level 2 of 5
Artificial General Intelligence
A proposed form of AI with broad capability across many kinds of tasks.
There is no universally accepted operational test for this label.
Example
Researchers discuss systems that could transfer competence across unfamiliar domains.
Listen to the definition and example
Audio transcript
Artificial General Intelligence.
A proposed form of AI with broad capability across many kinds of tasks.
There is no universally accepted operational test for this label.
For example: Researchers discuss systems that could transfer competence across unfamiliar domains.
Explore this concept
Why it matters
These distinctions help you separate a system’s task, representation, and reasoningmethod.
Artificial Intelligence.
Computing systems designed to perform tasks associated with human intelligence.
An AI system may recognize patterns, plan actions, or generate language without thinking like a person.
For example: A program identifies objects in a photo.
Explore this concept
Why it matters
These distinctions help you separate a system’s task, representation, and reasoningmethod.
Try answering before looking back at the definition.
Transformers · Level 3 of 5
Attention
A mechanism that weights information from candidate sources based on relevance.
Weights are computed from representations and need not be faithful explanations.
Example
A decoder emphasizes selected input positions when producing a token.
Listen to the definition and example
Audio transcript
Attention.
A mechanism that weights information from candidate sources based on relevance.
Weights are computed from representations and need not be faithful explanations.
For example: A decoder emphasizes selected input positions when producing a token.
Explore this concept
A useful analogy
A learned mixing desk that assigns different volumes to available information.
Why it matters
This helps you trace how token representations exchange and transform information.
“Remove my test data” clears all AI 101 and ML 101 scores, reviews, project checklists and rating selections, plus saved vocabulary and recent history in this browser. Test entries cannot be separated from other learning progress. Cookie preferences, security protections, submitted feedback, contact emails and past analytics are kept.