Out-of-Distribution (OOD) Detection

OOD detection fundamentally operates by modeling the probability density of the training distribution. A sample is flagged as OOD if it resides in a low-density region of the learned manifold.

• Parametric methods: Fit a Gaussian Mixture Model to learned embeddings.
• Non-parametric methods: Use k-Nearest Neighbors distance in the feature space.
• Deep generative models: Use the likelihood of a VAE or normalizing flow as a score, though these can paradoxically assign high likelihood to OOD inputs.

Energy-based models align with OOD detection by assigning a scalar energy score that is low for in-distribution data and high for OOD data. The Helmholtz free energy is computed directly from the logits.

• Energy score formula: E(x) = -T * log(sum(exp(f_i(x)/T)))
• Advantage: Energy scores are less susceptible to the overconfidence issue of softmax probabilities.
• Inference: No architectural changes are needed; the score is derived from the final classification layer.

The behavior of gradients with respect to model parameters provides a strong signal for OOD detection. In-distribution inputs produce coherent, stable gradients, while OOD inputs cause erratic gradient patterns.

• GradNorm: Uses the vector norm of the KL divergence gradient as a score.
• ODIN: Perturbs inputs with a small gradient step to increase the softmax temperature gap between ID and OOD samples.
• Feature: This method requires a backward pass, adding computational overhead at inference time.

Distance-based methods operate on the assumption that OOD samples are far from class prototypes or training samples in a learned embedding space.

• Mahalanobis distance: Computes the class-conditional Gaussian distance in a feature layer, requiring a covariance matrix per class.
• Deep Nearest Neighbors: Uses the distance to the k-th nearest training sample in the penultimate layer as the OOD score.
• Feature: These methods are simple, interpretable, and do not require retraining the classifier.

Bayesian frameworks separate predictive uncertainty into aleatoric and epistemic components. Epistemic uncertainty is the key signal for OOD detection.

• Epistemic uncertainty: High when the model lacks knowledge, characteristic of OOD regions. It is reducible with more data.
• Aleatoric uncertainty: High for noisy or ambiguous in-distribution data. It is irreducible.
• Mutual information: I(y;θ|x) quantifies epistemic uncertainty by measuring the disagreement between posterior samples in a Bayesian ensemble.

Unconstrained logit magnitudes can cause softmax to produce arbitrarily high confidence even for OOD inputs. Logit normalization mitigates this by bounding the logit vector.

• L2 normalization: Constrains ||f(x)||_2 = 1, preventing logit explosion.
• Cosine similarity: Replaces dot-product logits with cosine similarity, inherently normalizing the score.
• Effect: Normalization flattens the softmax distribution for OOD inputs, creating a clearer separation from ID confidence levels.

A probabilistic framework that learns an energy function mapping inputs to a single scalar value. The core principle: in-distribution data receives low energy, while OOD samples are assigned high energy. Unlike softmax-based classifiers that can be overconfident on OOD inputs, EBMs provide a density-aligned confidence score that naturally separates known from unknown distributions. Training often uses contrastive divergence or noise-contrastive estimation to shape the energy landscape, creating explicit low-energy valleys around training data.

An inference paradigm where a model is empowered with a reject option—the ability to abstain from prediction when confidence falls below a calibrated threshold. This directly operationalizes OOD detection by defining a coverage-confidence trade-off:

• Coverage: The fraction of inputs the model predicts on
• Risk: The error rate on accepted predictions
• Abstention: Inputs rejected due to low confidence The goal is to maximize coverage while maintaining a target risk level, often evaluated via the Risk-Coverage Curve and summarized by the Area Under the Risk-Coverage curve (AURC).

The reducible component of predictive uncertainty arising from the model's lack of knowledge. Unlike aleatoric uncertainty (inherent data noise), epistemic uncertainty is high in OOD regions where training data is sparse. Key characteristics:

• Decreases with more representative training data
• Increases far from the training manifold
• Captured by Bayesian methods like Deep Ensembles and Monte Carlo Dropout
• Essential for distinguishing between 'I don't know because the data is noisy' vs. 'I don't know because I've never seen this before' OOD detection fundamentally relies on identifying inputs that trigger high epistemic uncertainty.

A distribution-free, model-agnostic framework that wraps any predictor to produce prediction sets with finite-sample, marginal coverage guarantees. For OOD detection, conformal methods can construct calibrated rejection regions without assuming any specific data distribution. The process:

• Define a nonconformity score (e.g., 1 minus softmax confidence)
• Calibrate on a held-out set to determine a threshold
• At test time, include or reject based on the calibrated threshold Provides rigorous statistical guarantees: 'The true label will be in the prediction set with probability ≥ 1-α.'

The primary metric for evaluating how well a model's confidence scores align with its actual accuracy. ECE partitions predictions into M equal-width bins by confidence and computes the weighted average of the absolute difference between accuracy and confidence in each bin:

• ECE ≈ 0: Perfect calibration—70% confidence means 70% accuracy
• Overconfidence: Confidence exceeds accuracy (common in OOD inputs)
• Underconfidence: Accuracy exceeds confidence While ECE measures in-distribution calibration, poor ECE often correlates with poor OOD detection, as miscalibrated models produce unreliable confidence scores for rejection decisions.

A practical method for uncertainty quantification that trains multiple neural networks with different random initializations and aggregates their predictive distributions. For OOD detection:

• Ensemble disagreement signals epistemic uncertainty
• The variance across ensemble predictions is high for OOD inputs
• The mean confidence tends to be lower and better calibrated
• Provides a mixture-model predictive distribution that naturally separates aleatoric from epistemic uncertainty Unlike single-model methods, ensembles don't require architectural changes—just training M independent copies and averaging their softmax outputs at inference.