Out-of-Distribution (OOD) Detection

These techniques leverage the final-layer softmax probabilities of a classifier. The core assumption is that in-distribution (ID) samples will have high, confident predictions for a known class, while OOD samples will yield low, uniform probabilities.

• Maximum Softmax Probability (MSP): The simplest approach, using the highest predicted class probability as a confidence score. Low max probability indicates OOD.
• ODIN (Out-of-Distribution detector for Neural networks): Enhances MSP by using temperature scaling and adding small input perturbations to further separate ID and OOD score distributions.

Limitation: Modern neural networks are often overconfident, producing high softmax scores even for clearly OOD inputs, reducing the reliability of these methods.

These methods measure the similarity or distance of a new input's representation to known ID data within a learned feature space (e.g., the penultimate layer of a neural network).

• Mahalanobis Distance: Calculates the distance of a test sample's features to the closest class-conditional Gaussian distribution fitted on training data. Larger distances indicate OOD.
• k-Nearest Neighbors (k-NN): Uses the distance to the k-th nearest neighbor in the training set's feature space. OOD samples are expected to have larger neighbor distances.
• Cosine Similarity: Compares the angular similarity of the feature vector to prototype vectors or centroids of ID classes.

These methods are often more robust than softmax-based approaches as they operate on the richer feature space.

This family of techniques explicitly models the probability distribution of the training data in the feature space. OOD detection is performed by thresholding the estimated likelihood or density.

• Normalizing Flows: A class of generative models that learn an invertible transformation to map complex data distributions to a simple base distribution (e.g., Gaussian). The likelihood under the model can be computed exactly.

• Energy-Based Models (EBMs): Associate a scalar energy to each input configuration. ID data points are assigned lower energy. The energy score can be used directly for OOD detection: E(x) = -logsumexp(f(x)), where f(x) are the logits.

Challenge: These models can be complex to train and may assign higher likelihood to certain OOD samples—a phenomenon known as likelihood paradox.

This supervised approach trains the model explicitly to distinguish ID data from examples of what OOD data might look like, using an auxiliary outlier dataset.

• Outlier Exposure (OE): During training, the model is exposed to a diverse but unrelated auxiliary OOD dataset. The objective is modified to encourage lower confidence (e.g., uniform softmax distribution) on these auxiliary OOD examples while maintaining high confidence on ID data.

• Key Consideration: Performance is highly dependent on the choice of auxiliary dataset. The model learns the heuristic "anything that looks like these outliers is OOD," which may not generalize to all possible OOD inputs.

This method bridges the gap between pure unsupervised detection and having actual OOD labels.

These techniques analyze the model's internal behavior, such as its sensitivity to input changes or its predictive uncertainty, to infer OOD status.

• Gradient-Based Scoring: Examines the magnitude or pattern of gradients backpropagated through the network. OOD samples may produce different gradient signals compared to ID samples.

• Bayesian Neural Networks (BNNs) & Monte Carlo Dropout: Instead of a single point estimate, these methods produce a distribution over model parameters or predictions. OOD samples typically lead to higher predictive uncertainty (e.g., variance across multiple stochastic forward passes).

• Deep Ensembles: Trains multiple models with different initializations. Disagreement (predictive variance) among the ensemble members is higher for OOD inputs.

Modern, large-scale foundation models (e.g., CLIP, large vision models) offer a powerful, zero-shot approach to OOD detection without task-specific training.

• CLIP-based Detection: For vision, the cosine similarity between an image's embedding and a set of text embeddings (e.g., "a photo of a [class]") provides a confidence score. Low max similarity across classes can indicate OOD.

• Zero-Shot Confidence Scores: The inherent calibration and broad knowledge of foundation models can be harnessed. The model's internal scoring for its own generations or classifications can be thresholded.

• Advantage: Eliminates the need to train a dedicated OOD detector, leveraging the model's vast pre-existing knowledge of the visual or linguistic world.

Data drift, often synonymous with covariate shift, occurs when the statistical distribution of the input features (X) changes between the model's training environment and its production environment, while the relationship P(Y|X) remains constant. It is a primary cause of model performance decay.

• Detection Method: Compare feature distributions (e.g., using PSI, KL Divergence) between a baseline distribution (training set) and current inference data.
• Example: An e-commerce model trained on user data from 2022 sees a shift in 2024 as a new age demographic becomes the primary user base, altering feature distributions for 'age' and 'browsing time'.

Concept drift is the change in the statistical relationship between the input features (X) and the target variable (Y) over time. The mapping the model learned becomes incorrect, even if the input distribution is stable.

• Key Difference from Data Drift: The fundamental P(Y|X) changes.
• Example: A credit fraud model's patterns become obsolete because fraudsters develop new techniques, making old feature correlations (e.g., transaction amount, location) less predictive.
• Detection: Requires ground truth labels to monitor performance metrics like accuracy or F1-score for degradation, or uses specialized algorithms like the Page-Hinkley Test on prediction errors.

Model Performance Monitoring (MPM) is the overarching practice of tracking a deployed model's key business and accuracy metrics to detect degradation. It is the operational layer that uses drift detection as a leading indicator.

• Core Metrics: Accuracy, precision, recall, F1-score, custom business KPIs.
• Relationship to OOD: A sustained influx of OOD data often leads to a measurable drop in MPM metrics. MPM provides the 'ground truth' signal that drift detection algorithms try to predict.
• Implementation: Involves setting SLOs/SLIs for AI, A/B testing frameworks, and experiment tracking systems to compare new model versions against baselines.

These are mathematical functions that quantify the difference between two probability distributions, forming the computational backbone of unsupervised drift detection.

• Population Stability Index (PSI): A robust, interpretable metric for univariate or multivariate drift. Values > 0.1 suggest minor drift, > 0.25 indicates major shift.
• Kullback-Leibler (KL) Divergence: Measures information loss when one distribution is used to approximate another. Asymmetric and sensitive to small probability differences.
• Wasserstein Distance (Earth Mover's Distance): Measures the minimum 'cost' to transform one distribution into another. More geometrically intuitive and works well with sparse or high-dimensional data.
• Chi-Squared Test: A hypothesis test for categorical data to determine if observed frequency counts differ from expected (baseline) counts.

This distinction defines the temporal paradigm of drift detection, impacting system latency and architecture.

• Online Drift Detection: Analyzes a continuous data stream in real-time, often using a sliding window or algorithms like ADWIN (Adaptive Windowing). Aims to minimize detection delay for sudden drift.
• Batch Drift Detection: Periodically analyzes accumulated data (e.g., hourly, daily). Compares statistics of a recent batch against the baseline distribution. More suitable for detecting gradual drift and environments where ground truth labels arrive with delay.
• Trade-off: Online detection offers speed but higher computational cost and potential for noise; batch detection is more stable but introduces latency.

The set of strategies invoked once drift is detected to restore model performance, closing the MLOps feedback loop.

• Automated Retraining Pipeline: An orchestrated workflow triggered by drift alerts or performance thresholds. It fetches new labeled data, retrains the model (potentially using Continuous Model Learning techniques), validates it, and redeploys.
• Drift Adaptation: Broader strategies including:
  • Online Learning: Incrementally updating model weights with new data.
  • Ensemble Methods: Weighting newer models more heavily.
  • Contextual Bandits: Dynamically selecting the best model from a pool.
• Root Cause Analysis (RCA) for Drift: The investigative process to determine if drift is due to data pipeline issues (training-serving skew), genuine environmental change, or a faulty detection (false positive).