Learn AI. Clear definitions for the concepts you’ll meet along the way. Start with the idea, then see it in action.
Four constrained numbers used to represent a 3D rotation. The component order must be declared; opposite signs represent the same orientation. A quaternion does not describe translation or establish feasible motion.
Explore: Represent motion in three dimensionsAn origin and directed axes used to describe a position or orientation. The same physical point has different coordinates in different frames; transformations require named source/destination frames and consistent units.
Explore: Describe where the robot isCalculating body or tool position and orientation from a robot’s joint configuration and geometry. It does not by itself solve for joints, test collisions or establish physically feasible motion.
Explore: Describe where the robot isA declared comparison reports data, supervision, interactions, optimizer work, checkpoint selection and every seed or failure. Frozen evaluation keeps task success separate from a potentially misleading reward proxy.
Explore: Evaluate learning fairlyLearning an action policy from a fixed log without collecting additional experience during fitting. Repeated updates do not fill missing state/action coverage, and conservative estimates do not create unobserved outcomes.
Explore: Learn without new interactionsA continuous-action learning approach that combines a stochastic actor, replay, critics and an entropy-weighted objective. Entropy temperature trades off predicted return and action-distribution spread; it is not a safety guarantee.
Explore: Trade off reward and exploration with entropyA policy surrogate uses the smaller of a probability-ratio-weighted advantage and its clipped-ratio version. It removes some incentives for large changes, but does not enforce a hard probability or KL limit.
Explore: Limit a policy updateA learner with a policy (actor) and a return estimator (critic). A critic can supply an advantage estimate for policy updates, but inaccurate bootstrapped values can bias that signal.
Explore: Use a critic to guide learningA derivative of expected return with respect to policy parameters. A sampled estimator can weight log-probability derivatives by returns minus an action-independent baseline; individual updates can be noisy.
Explore: Increase the probability of useful actionsUsing data collected by a behavior policy to learn about or improve a different target policy. Adequate coverage, the estimator’s assumptions and the actual logged collection probabilities matter.
Explore: Know which policy produced the dataSupervised learning that fits a policy to demonstrated action targets. Agreement on demonstrations does not establish recovery skill or successful complete attempts.
Explore: Copy a demonstrated actionAn agent receives observations that may omit or obscure environment state. Different states can share an observation; history or a justified belief can help, without guaranteeing exact state recovery.
Explore: What can the agent observe?A consistency condition relating a state value to immediate reward and discounted expected future value. Policy evaluation follows fixed actions; the optimality equation maximizes over actions.
Explore: Solve the Bellman equationsA common cause influences both an exposure or assignment and its outcome, so an observed association may differ from the effect of intervening. Adjustment requires appropriate causal assumptions, measurements and overlap.
Explore: Ask what would change under interventionAn outcome under an alternative intervention for the same unit. Population observational and intervention rates may leave the pairing of individual potential outcomes undetermined without additional structural assumptions.
Explore: Ask what would change under interventionChoose the action with the best worst-case opponent response in a zero-sum game. Chance outcomes still use their probabilities; guarantees depend on the model and utility.
Explore: Reason about an opponentChoose the action with greatest expected utility using supplied probabilities for uncertain outcomes or opponent responses. Incorrect probabilities can produce poor decisions even with exact search.
Explore: Reason about an opponentAn approximation of a hidden-state belief using samples that move through a transition model, receive observation likelihood weights and may be resampled. Finite-sample error and lost diversity are separate from model error.
Explore: Approximate a belief with particlesThe reciprocal of the sum of squared normalized particle weights. Measures weight concentration; it is not a count of unique states, independent observations or correct hypotheses.
Explore: Approximate a belief with particlesEstimating the current hidden-state distribution using only observations up to the current time. Prediction advances the state model; correction incorporates the current reading. Exact filtering does not guarantee the model matches reality.
Explore: Infer state across timeA directed acyclic graph and local conditional probability tables that factor a joint distribution and encode conditional-independence assumptions. The arrows alone do not establish causation.
Explore: Factor a probabilistic modelIn a specified causal model, replacing a variable’s ordinary assignment mechanism with a controlled setting while leaving other mechanisms unchanged. This differs from observing its naturally occurring value.
Explore: Factor a probabilistic modelA probability assigned before the evidence currently being considered. It should describe the relevant population or assumptions, not a convenient unrelated group.
Explore: Update beliefs from evidenceHow probable specified evidence is under a proposed state or model. Probability of an alarm given a fault differs from probability of a fault given an alarm.
Explore: Update beliefs from evidenceA probability conditioned on observed evidence. For finite events, conditioning requires a positive probability of that evidence under the model.
Explore: Update beliefs from evidenceAfter a specified condition is known, learning one variable supplies no further information about the other. Conditional independence does not generally imply independence in the mixed population.
Explore: Update beliefs from evidenceChoose one value for each variable from its allowed domain so every stated relationship holds. Feasibility alone does not optimize preferences or prove the requirements describe reality.
Explore: Solve a constraint problemFor each directed binary relationship, every remaining value has at least one compatible value in the neighboring domain. These separate supports need not combine into a global solution.
Explore: Solve a constraint problemExplore a tentative choice, then undo its branch-specific assignments and deductions when it fails before trying another choice.
Explore: Solve a constraint problemEstimate preferences using patterns of observed feedback across users and items. Missing entries are not automatically negative ratings; learned factors need evidence and cold-start handling.
Explore: Recommend from incomplete feedbackA user or item lacks the history a method needs. Content, onboarding or explicit fallbacks may help; absent feedback is not a verified preference.
Explore: Recommend from incomplete feedbackRecorded feedback reflects what was exposed, chosen and logged. It need not represent preferences over all available options, and recommendations can reinforce that selection.
Explore: Recommend from incomplete feedbackA positive prediction or flag for a case whose reference label is negative. In anomaly review this is a false alarm on a confirmed normal case.
Explore: Find unusual examplesTrue positives divided by all positive predictions: TP / (TP + FP). It is undefined when no positives are predicted, and depends on the population mix.
Explore: Find unusual examplesTrue positives divided by all actual positives: TP / (TP + FN). Also called the true-positive or detection rate; undefined if no positive reference cases are available.
Explore: Find unusual examplesScoring unusualness relative to a fitted reference. An anomaly flag is a reason to investigate, not a confirmed problem; coverage, thresholds and base rates affect interpretation.
Explore: Find unusual examplesAverage squared deviation from a mean, describing numerical spread. These cohort examples divide by the number of observations; retained variance is not task accuracy or semantic importance.
Explore: Find structure without labelsA clustering algorithm alternating nearest-center assignments with moves to group means under squared Euclidean distance. Initialization, scaling and k affect results; groups are not verified labels.
Explore: Find structure without labelsA linear representation using directions of large centered variance. Keeping fewer directions can discard task-relevant distinctions even when retained variance is high.
Explore: Find structure without labelsSubtract a fitted coordinate mean and divide by its fitted spread. Fit statistics on the intended development/training cohort and reuse them; constant coordinates need explicit zero-spread handling.
Explore: Package a model as a componentAn additive model fitted in sequence, with each new learner approximating a direction that reduces the current loss. For squared-error regression, the next tree fits observed-minus-predicted residuals.
Explore: Correct residual errors with boostingIn boosting, a learning-rate multiplier applied to each new learner before adding it to the existing prediction. Smaller corrections often need more stages; they do not guarantee better generalization.
Explore: Correct residual errors with boostingAn ensemble of trees trained on resampled data with random subsets of features considered at each split.
Explore: Combine many imperfect treesA sample drawn with replacement, so an original row can appear repeatedly or be omitted.
Explore: Combine many imperfect treesEvaluation of each training row using only trees whose bootstrap samples omitted that row. Report coverage and treat repeated tuning as development.
Explore: Combine many imperfect treesA model that applies a sigmoid to a weighted input sum to estimate a binary class probability. Training fits probabilities; a separate threshold chooses class decisions.
Explore: Predict a class probabilityA reference method used to judge whether another approach adds value. Fit a mean or majority predictor from training data and compare candidates on the same separate examples.
Explore: Start with a simple baselineA method that predicts from nearby training examples under a defined distance. Classification may use a majority vote; scale, neighborhood size and tie rules affect the result.
Explore: Start with a simple baselineA predictor that follows learned input tests to a leaf. Training chooses splits; leaf predictions summarize training examples. More depth increases flexibility, not guaranteed generalization.
Explore: Start with a simple baselineA numerical predictor formed from an intercept and weighted inputs. Least squares fits its coefficients by minimizing training squared error; an observed coefficient is not automatically causal.
Explore: Fit a line and explain its errorsObserved value minus predicted value in these regression lessons. A signed prediction error, distinct from a neural-network residual connection.
Explore: Fit a line and explain its errorsThe average of squared prediction residuals. Its units are squared output units, larger errors contribute disproportionately, and an empty-sample mean is undefined.
Explore: Fit a line and explain its errorsThe average absolute size of prediction residuals, in the original output units. It does not allow positive and negative errors to cancel.
Explore: Fit a line and explain its errorsA representation or computation used to describe or predict something. A learned model fits parameters from examples; a programmed rule specifies its behavior directly.
Explore: What should the machine predict?An estimated output for an input, such as a category or a number. It is not automatically an action, a verified fact or a guarantee.
Explore: What should the machine predict?Examples used to fit model parameters or preprocessing. Evaluation evidence must be separated according to the claim, including related groups or time periods when needed.
Explore: Examples are measurementsA feed-forward neural network with learned affine transformations and nonlinear hidden layers. Its inputs and outputs depend on the task representation.
Explore: Train and evaluate a small neural routerTasks and directed prerequisite relationships specifying which outputs must be available before other tasks start.
Explore: Coordinate dependencies and reusable skillsA longest chain of dependent tasks. Its total duration places a lower bound on completion time under the stated timing assumptions.
Explore: Coordinate dependencies and reusable skillsA temporally extended action defined by its initiation conditions, internal policy, and termination rule. It can be programmed or learned.
Explore: Coordinate dependencies and reusable skillsComparing a proposed change’s base version with the current artifact version to detect stale writes before accepting them.
Explore: Coordinate dependencies and reusable skillsRepeating an operation has the same relevant effect as performing it once. Stable operation identities and saved receipts can prevent duplicate effects.
Explore: Coordinate dependencies and reusable skillsDetermining how actions or components contributed to an outcome, including interactions that make local activity a poor measure of joint success.
Explore: Coordinate dependencies and reusable skillsInformation retained for later use, in current working state, explicit records, reusable procedures, or learned parameters. Retrieval and correctness require separate checks.
Explore: Remember selectively and retrieve evidenceDeciding which computation or evidence-gathering action to perform, considering its expected benefit, cost, and stopping conditions.
Explore: Decide whether another computation is worth itThe expected improvement in a decision made after seeing evidence. Net value subtracts the cost of obtaining that evidence.
Explore: Decide whether another computation is worth itAgreement between predicted probabilities and observed frequencies over comparable cases. A calibrated probability is not a guarantee for an individual case.
Explore: Predict a class probabilityA controller’s choices change which tasks and outcomes appear in each specialist’s records, making raw comparisons potentially misleading.
Explore: Learn from routing evidence without fooling yourselfA decision process that uses the current task’s features to choose an action and learns from its observed outcome, without directly optimizing a long sequence of state changes.
Explore: Learn from routing evidence without fooling yourselfTask-relevant information kept available for current decisions, such as the active goal, constraints, and recent observations.
Explore: Share the right information at the right timeA shared place where selected goals, observations, and memory references are made available to multiple components.
Explore: Share the right information at the right timeDeciding which information or computation gets priority at the system level. This is distinct from a Transformer’s internal attention calculation.
Explore: Share the right information at the right timeA feedback process that responds near the component doing the work, while higher-level planning supplies objectives and handles exceptions.
Explore: Share the right information at the right timeRecords of particular events or experiences, with their context, actions, and outcomes.
Explore: Remember selectively and retrieve evidenceRetained facts, concepts, and relationships, with applicable scope and evidence when represented as explicit records.
Explore: Remember selectively and retrieve evidenceReusable ways to carry out tasks, represented as stored procedures, controllers, or learned skills.
Explore: Remember selectively and retrieve evidenceFinding or weighting stored information using a query’s relationship to memory keys. Similarity alone does not verify the result.
Explore: Remember selectively and retrieve evidenceTransforming selected experience into more durable or reusable knowledge or skills while preserving scope and checking counterexamples.
Explore: Remember selectively and retrieve evidenceA shared structured form with explicit fields, entities, and relations that connects different processing stages.
Explore: Learn a shared representation contractLearning representations by favoring matched examples over contrasting alternatives. The chosen comparisons determine which distinctions the objective rewards.
Explore: Learn a shared representation contractA neural model that processes nodes and their relationships, often by passing and aggregating messages along graph edges before learned node updates.
Explore: Learn a shared representation contractOptimizing multiple connected components with a shared objective. Gradients can coordinate their representations while also introducing coupling and regressions.
Explore: Choose where a connected system learnsA declared requirement a candidate version must satisfy before promotion, such as a critical accuracy floor or resource ceiling.
Explore: Test a proposed system improvementRestoring a recorded previous version after a change. It recovers an earlier state but does not prove that state meets every current requirement.
Explore: Test a proposed system improvementAdapting choices to repeated benchmark feedback so the score improves without a corresponding improvement on new tasks.
Explore: Find out whether improvement generalizesThe average squared difference between a predicted probability and a binary outcome. Lower is better; it measures probability error.
Explore: Find out whether improvement generalizesFurther training some or all existing model parameters for a chosen task or dataset, with separate evaluation of gains and retained behavior.
Explore: Choose where a connected system learnsAgreement between connected components about the meaning and use of exchanged information. Matching shapes or schemas alone does not establish it.
Explore: Choose where a connected system learnsA structured request to an externally implemented operation. A proposal only becomes an action if the execution runtime accepts and runs it.
Explore: Call a tool, then check the evidenceThe component that enforces allowed actions, validates arguments, runs operations, and returns observations independently of a model’s proposal.
Explore: Call a tool, then check the evidenceWhere evidence came from: its source, request, version, and relevant timing. Useful for deciding whether an observation applies to the current question.
Explore: Call a tool, then check the evidenceThe agreed form and interpretation of information passed between components, including fields, shapes, units, and meaning.
Explore: Help specialists understand each otherA transformation connecting representations expected by different components. It can be programmed or learned and must be evaluated for compatibility.
Explore: Help specialists understand each otherAn architecture routing inputs through selected expert blocks within a network. This differs from choosing independent full models or tools for subtasks.
Explore: Help specialists understand each otherA representation with events as nodes and relations such as before or after as edges, often retaining sources and timestamps.
Explore: Help specialists understand each otherA component that proposes examples, answers, or action sequences. Its proposal is not independent evidence of correctness.
Explore: Build a trustworthy practice loopA rule or strategy for selecting which tasks or examples to practice, often based on difficulty or current performance.
Explore: Build a trustworthy practice loopTraining a model to imitate targets from a teacher or a more expensive process such as search. Teacher errors and coverage limits can transfer too.
Explore: Build a trustworthy practice loopSubstantial loss of older capabilities after parameter updates for new tasks.
Explore: Learn something new without losing the oldReusing old examples during new training so earlier behavior still influences parameter updates.
Explore: Learn something new without losing the oldLearning an initialization, representation, or strategy that helps adaptation across tasks, evaluated after a limited adaptation budget.
Explore: Learn a better way to adaptChanging behavior for a new task using additional evidence. Parameters, temporary context, or another retained state may be involved; specify which.
Explore: Learn a better way to adaptExamples supplied for task-specific adaptation. Separate query examples measure the adapted result.
Explore: Learn a better way to adaptExamples used to evaluate an adapted model. During meta-training, their loss can train shared meta-parameters through the outer objective.
Explore: Learn a better way to adaptChanging behavior in response to examples in the current input context, ordinarily without updating stored model parameters.
Explore: Learn a better way to adaptExploring explicit alternatives, such as action sequences, while tracking which possibilities remain to be checked.
Explore: Reason with procedures and searchA specified procedure for transforming inputs into results. Its guarantees depend on assumptions about inputs, rules, and resources.
Explore: Reason with procedures and searchAn estimate or rule that helps choose promising options. A useful clue is not a proof of correctness.
Explore: Reason with procedures and searchA model that predicts how an environment changes, conditioned on its state and an action. Prediction errors can compound during planning.
Explore: Predict what an action will changePlanning ahead, executing the first action, observing the actual result, and planning again from the updated state.
Explore: Predict what an action will changeInformation used to describe a situation for prediction or decision-making. Missing relevant variables can make different situations look identical.
Explore: Predict what an action will changeReusing an update operation over successive internal states. Repetition can be programmed or learned and does not guarantee progress.
Explore: Refine an idea over several passesAn internal representation used by a model, often a vector. It need not be written as words or directly match observable quantities.
Explore: Refine an idea over several passesApproaching a stable value or state under an update process. Stability alone does not establish task correctness.
Explore: Refine an idea over several passesA controlled comparison that removes or changes a component to investigate its contribution, including interactions and resource costs.
Explore: Test the parts of a planning systemA numerical signal supplied after an action. It specifies what learning is encouraged to increase, which may differ from the designer’s intended task.
Explore: Explore before you decideAccumulated rewards over time, often discounted so that more distant rewards contribute less.
Explore: Learn from delayed rewardsChoosing actions to collect information about alternatives. Exploitation uses current estimates to select a promising action.
Explore: Explore before you decideExpected return associated with taking an action in a state, under specified later behavior. A learned estimate can be wrong.
Explore: Explore before you decideExpected return from a state when following a particular policy. In the one-step bandit, it is the policy-weighted average of action values.
Explore: Explore before you decideThe difference Q − V: how much better an action is estimated to be than the policy’s state-value baseline.
Explore: Explore before you decideAn action-value learning method that moves an estimate toward observed reward plus discounted maximum next-state value; terminal outcomes have no future contribution.
Explore: Learn from delayed rewardsObtaining a high reward by exploiting the reward rule while failing the intended task.
Explore: Learn from delayed rewardsGenerating training experience by having an agent or copies of its policy play a game. Reliable outcomes and independent evaluation still matter.
Explore: Learn from verified self-playA procedure that checks a claim or result against specified criteria, such as exact game rules. Its reliability is limited by what those checks actually establish.
Explore: Learn from verified self-playA smooth nonlinear activation between −1 and 1. Its slope at an output a is 1 − a², which is used during backpropagation.
Explore: the classifier projectA table that separates predictions by their observed and predicted classes, revealing different kinds of classification mistakes.
Explore: the classifier projectA short account of a model’s purpose, data, evaluation, limitations, and appropriate use. It connects performance claims to evidence.
Explore: the classifier projectA method that discourages some solutions during learning, such as penalizing large weights. It can reduce overfitting but must be evaluated rather than assumed helpful.
Explore: the classifier projectA count of floating-point arithmetic operations. FLOPs per second is a processing rate; neither quantity is memory capacity.
Explore: Spend a limited compute budgetA stored number adjusted during learning, such as a weight or bias. Parameter count alone does not measure model quality.
Explore: Spend a limited compute budgetThe rate at which bytes move to or from memory. Moving weights and intermediate results can limit computation speed.
Explore: Spend a limited compute budgetActivating different parts of a computation for different inputs, such as routing tokens to selected experts. Inactive parameters can still occupy storage.
Explore: Spend a limited compute budgetAn empirical relationship between resources such as data or model size and measured behavior such as loss. Its reliability is limited to the evidence and conditions supporting it.
Explore: Spend a limited compute budgetAn ordered list of numbers. Matching coordinates are multiplied and added in a dot product; the order of coordinates matters.
Explore: Give words a place in spaceA comparison of directions: the dot product divided by the two vector lengths. It is undefined for a zero vector.
Explore: Give words a place in spaceA representation in which combinations of coordinates carry useful distinctions, rather than assigning each concept to one isolated neuron.
Explore: Give words a place in spaceAn attention query compares with keys to determine mixture weights. Those weights combine value vectors into an output.
Explore: Compare, normalize, combineA transformation from numerical scores into nonnegative weights that sum to one, using exponentials and normalization.
Explore: Compare, normalize, combineAn information boundary that allows each decoder position to attend only to itself and earlier positions.
Explore: Compare, normalize, combineA neural architecture combining attention with feed-forward transformations, residual connections, and normalization. Decoder-only models can predict the next token.
Explore: Follow a token through a TransformerA token is an entry in a model’s vocabulary. A tokenizer converts text into token IDs, often splitting words into smaller pieces.
Explore: Follow a token through a TransformerAn addition that combines a block’s input with a transformed result, preserving a direct path for information and gradients.
Explore: Follow a token through a TransformerRescaling numerical values using a selected rule. Layer normalization centers and rescales coordinates within each token representation.
Explore: Follow a token through a TransformerA raw score before conversion to probabilities. Vocabulary logits score possible next tokens.
Explore: Follow a token through a TransformerStored attention keys and values from earlier positions, reused during causal generation to avoid recomputing them. It costs memory and does not update weights.
Explore: Follow a token through a TransformerFitting training-specific noise or details that do not carry over to new examples.
Explore: Learn the pattern, not the examplesA learning process’s preferences among possible patterns. Architecture, optimization, regularization, and data all contribute.
Explore: Learn the pattern, not the examplesExamples excluded from parameter fitting and used to choose settings. A separate test set evaluates the final frozen choice.
Explore: Learn the pattern, not the examplesA pattern in inputs or activations that a computation responds to. A feature may be distributed across many neurons.
Explore: What does this neuron notice?A smooth function that maps a numerical score to a value between 0 and 1. Interpreting it as a probability does not guarantee calibration.
Explore: What does this neuron notice?A prediction loss that penalizes assigning little probability to the correct outcome.
Explore: What does this neuron notice?Using a predictive signal that works in the observed data but does not represent the intended task, such as a background instead of an animal’s shape.
Explore: Catch a model taking a shortcutA change in the conditions or patterns of the data, which can make previous performance estimates unreliable.
Explore: Catch a model taking a shortcutDeliberately changing a factor to investigate its effects, with explicit assumptions about what else stays fixed.
Explore: Catch a model taking a shortcutA calculation that multiplies inputs by weights, adds a bias, and may apply an activation function.
Explore: Your first neuronA learned multiplier. In our price example, it controls how much the prediction increases for each extra item.
Explore: Your first neuronA learned offset added to the weighted input. It lets the model predict a nonzero value when the input is zero.
Explore: Your first neuronA number measuring prediction error. Our lab averages the squared errors over four examples.
Explore: Follow the slope of a mistakeThe local sensitivity of loss to each parameter. Gradient descent takes a small step in the opposite direction.
Explore: Follow the slope of a mistakeA layer’s output after its transformation. A nonlinear activation lets a network express more than a straight line.
Explore: Make a problem easier to seeA method for calculating gradients through connected computations. An optimizer then uses those gradients to update weights.
Explore: Send the error backwardHow well a model performs on examples or conditions it did not use for training.
Explore: Learn the pattern, not the examplesAn ordered list of numbers representing something such as a word, image, or game state.
Explore: Give words a place in spaceA mechanism that compares queries and keys to decide how to combine information from value vectors.
Explore: Compare, normalize, combineA rule or learned function for choosing an action, sometimes by assigning probabilities.
Explore: Explore before you decideA controller that selects and coordinates tools, models, and actions over time.
Explore: Share the right information at the right time