Distributional Shift

A change in the distribution of the input features P(X) while the conditional relationship P(Y|X) remains constant.

• Example: A self-driving car trained in sunny California deployed in snowy Norway. The input pixel distribution changes, but the rules of the road stay the same.
• Detection: Statistical two-sample tests (Kolmogorov-Smirnov) comparing feature means and variances between training and production data.
• Mitigation: Importance weighting and domain adaptation techniques that re-weight training samples to match the target distribution.

A change in the prior probability of the target variable P(Y) while the class-conditional distribution P(X|Y) remains fixed.

• Example: A disease screening model trained on a hospital population with 20% prevalence deployed to a general population with 2% prevalence.
• Consequence: The model's calibration breaks; predicted probabilities become systematically overconfident or underconfident.
• Mitigation: Adjusting output probabilities using the known base rate shift via Bayes' theorem or Platt scaling with a calibration set from the target domain.

A change in the fundamental relationship between inputs and outputs P(Y|X) over time, making the learned decision boundary obsolete.

• Example: A fraud detection model where fraudsters continuously adapt their tactics. The mapping from transaction features to 'fraudulent' changes.
• Types:
  • Sudden drift: Abrupt change due to an external shock (e.g., a new regulation).
  • Incremental drift: Gradual evolution (e.g., user preferences slowly changing).
  • Recurring drift: Cyclical patterns (e.g., seasonal buying behavior).
• Mitigation: Online learning with a forgetting factor, sliding window retraining, and drift detection algorithms like ADWIN.

The appearance of entirely novel classes or feature combinations in deployment that were absent from the training distribution.

• Example: An image classifier trained on 1,000 ImageNet classes encountering a completely new object category it has no representation for.
• Failure Mode: The model confidently misclassifies the unknown input into one of its known classes with high softmax probability.
• Mitigation: Out-of-distribution detection using energy-based models, Mahalanobis distance in feature space, or adding an explicit 'reject' class with outlier exposure training.

A change in the relative frequency of sub-groups within the data, causing a model that performs well on average to fail on underrepresented slices.

• Example: A speech recognition system trained on balanced dialect data deployed in a region where one dialect dominates, causing degraded performance on minority dialects.
• Diagnosis: Slice-based evaluation that disaggregates performance metrics across known sub-groups rather than relying on aggregate accuracy.
• Mitigation: Group distributionally robust optimization (Group DRO) that minimizes worst-case loss across subpopulations during training.

A distributional change driven purely by the passage of time, encompassing evolving cultural norms, language usage, and factual knowledge.

• Example: A language model trained on pre-2020 data that cannot answer questions about post-2020 events or uses outdated terminology.
• Temporal Degradation: Model performance decays monotonically as the gap between the training data cutoff and the current date widens.
• Mitigation: Continuous fine-tuning on temporally relevant data, retrieval-augmented generation to ground responses in current documents, and time-aware model architectures.

A specific type of distributional shift where the statistical relationship between input features and the target variable changes over time. Unlike simple data drift, concept drift alters the fundamental rules the model learned.

• Sudden Drift: An abrupt change, like a new fraud pattern emerging overnight.
• Incremental Drift: A gradual change, such as user preferences evolving over months.
• Recurring Drift: A cyclical change, like seasonal purchasing behaviors.

The task of identifying input data that is fundamentally different from the model's training distribution. This acts as a safety trigger, allowing the system to reject a prediction or fall back to a safe default instead of extrapolating wildly.

• Softmax Thresholding: Rejecting inputs with low maximum predicted probability.
• Mahalanobis Distance: Measuring the distance of a feature vector from the training data's class centroids.
• Energy-Based Models: Using a model's energy score to separate in-distribution from out-of-distribution samples.

A learning failure where an agent infers spurious correlations as causal relationships. When a non-causal feature that was predictive in training shifts in deployment, the policy becomes brittle and misgeneralizes.

• Example: A self-driving car learns to steer based on the presence of a lane-line crack, not the road edge. When the crack is repaved, the policy fails.
• Mitigation: Causal representation learning and domain randomization force the model to learn invariant, causal features.

The discrepancy between a simulated training environment and the physical world that causes policies trained in simulation to fail upon real-world deployment. This is a fundamental distributional shift for embodied agents.

• Visual Gap: Rendered textures differ from real-world lighting and materials.
• Physics Gap: Simulated dynamics don't perfectly model friction, mass, or contact forces.
• Domain Randomization: Varying simulation parameters (lighting, textures) extensively to force the policy to generalize to reality.

The property of an AI system's goal-directed behavior remaining consistent and correct under distributional shift or adversarial perturbation. A robust objective does not shatter when the environment changes.

• Adversarial Training: Exposing the agent to worst-case input perturbations during training.
• Quantile Risk Metrics: Optimizing for conditional value-at-risk (CVaR) instead of average performance to handle tail events.
• Invariant Risk Minimization: Learning representations that elicit the same optimal predictor across different training environments.

A state where an AI agent's internal world model or category system breaks down due to a fundamental shift in its environment or capabilities. The agent's ontology no longer carves reality at the joints.

• Example: A trading agent's model of 'market volatility' shatters when a new financial instrument fundamentally alters market microstructure.
• Response: Requires meta-learning to recognize when its own categories are failing and trigger a re-learning or safe shutdown process.