Glossary

ODIN

ODIN (Out-of-DIstribution detector for Neural networks) is a method that applies temperature scaling and small adversarial perturbations to inputs, amplifying the difference in SoftMax scores between in-distribution and out-of-distribution samples.

Get in touch Learn more

Finance professional using AI FP&A copilot on laptop, board presentation visible on screen, home office work session.

OUT-OF-DISTRIBUTION DETECTOR

What is ODIN?

ODIN is an out-of-distribution detector that applies temperature scaling and small adversarial perturbations to inputs, amplifying the difference in SoftMax scores between in-distribution and out-of-distribution samples.

ODIN (Out-of-DIstribution detector for Neural networks) is a post-hoc method that enhances a pre-trained neural network's ability to separate known in-distribution samples from unknown out-of-distribution inputs without retraining. It operates by first applying temperature scaling to soften the SoftMax probability distribution, then adding a small, calibrated adversarial perturbation to each input. This perturbation is calculated by taking the sign of the gradient of the temperature-scaled SoftMax score with respect to the input, effectively pushing in-distribution samples toward higher confidence and out-of-distribution samples toward lower confidence, widening the separability gap.

The core mechanism relies on the observation that in-distribution inputs respond differently to input pre-processing than out-of-distribution samples. By subtracting a weighted step in the direction that increases the maximum class score, ODIN causes the predicted probabilities for known classes to become more pronounced while forcing unknown inputs into a uniform, low-confidence state. A threshold on the maximum calibrated and perturbed SoftMax score then serves as the decision boundary for rejection. This lightweight, inference-time technique requires no auxiliary outlier dataset for tuning, making it a practical first-line defense against distributional shift in deployed classification systems.

OUT-OF-DISTRIBUTION DETECTOR

Key Features of ODIN

ODIN (Out-of-DIstribution detector for Neural networks) is a lightweight, post-hoc method that enhances any pre-trained neural network's ability to separate known from unknown inputs without retraining.

Temperature Scaling

Applies a temperature parameter T to the SoftMax function during inference, dividing logits to soften probability distributions. Higher temperatures amplify the difference between in-distribution and out-of-distribution SoftMax scores.

No model retraining required
Single scalar parameter tuned on validation data
Preserves original classification accuracy on known classes

T > 1

Typical Temperature Range

Input Preprocessing with Adversarial Perturbations

Adds a small, calibrated perturbation to each input sample before scoring. The perturbation is computed as the sign of the gradient of the SoftMax score with respect to the input, pushing in-distribution samples toward higher confidence while leaving out-of-distribution samples relatively unaffected.

Perturbation magnitude controlled by epsilon (ε)
Requires only a single backward pass
Amplifies the separability gap between known and unknown

Threshold-Based Detection

After temperature scaling and perturbation, ODIN computes the maximum SoftMax score for each input. Samples with scores below a calibrated threshold are flagged as out-of-distribution.

Simple binary decision rule
Threshold selected using a held-out validation set of OOD examples
Compatible with any pre-trained classifier without architectural changes

Hyperparameter Tuning Protocol

ODIN requires tuning two hyperparameters: temperature (T) and perturbation magnitude (ε). These are selected jointly on a validation set containing both in-distribution and out-of-distribution samples to maximize detection performance.

Grid search over T and ε values
Optimizes AUROC or TNR at 95% TPR
No access to test-time OOD data required

Compatibility with Pre-Trained Models

ODIN operates as a post-hoc wrapper around any neural network with a SoftMax output layer. It requires no modification to the model architecture, no access to training data, and no retraining of weights.

Works with CNNs, ResNets, DenseNets, and transformers
Applicable to image, signal, and text classification
Demonstrated on CIFAR-10, CIFAR-100, and ImageNet benchmarks

Performance Benchmarks

ODIN significantly outperforms the baseline maximum SoftMax probability detector. On standard benchmarks, it reduces the false positive rate at 95% true positive rate (FPR95) by up to 34% compared to uncalibrated SoftMax.

Evaluated against datasets like TinyImageNet, LSUN, and iSUN as OOD
Maintains near-identical in-distribution accuracy
Competitive with more complex methods requiring retraining

34%

FPR95 Reduction vs. Baseline

ODIN DETECTOR EXPLAINED

Frequently Asked Questions

Clear, technical answers to the most common questions about the ODIN out-of-distribution detector, its mechanisms, and its application in open-set signal recognition.

ODIN (Out-of-DIstribution detector for Neural networks) is a post-hoc, inference-time method that detects out-of-distribution (OOD) samples by applying temperature scaling and small adversarial perturbations to inputs, amplifying the difference in SoftMax scores between in-distribution and out-of-distribution data. Unlike methods requiring model retraining, ODIN operates on a pre-trained neural network without modifying its weights. The technique is hyperparameter-free in its core logic, requiring only a calibration set to tune the temperature and perturbation magnitude. For a modulation classifier, ODIN takes an IQ sample, adds a tiny perturbation calculated from the gradient of the SoftMax output, and then re-evaluates the prediction with a high temperature. If the recalibrated maximum class probability falls below a threshold, the signal is flagged as an unknown modulation scheme, making it a critical tool for open set signal recognition in dynamic spectrum environments.

Enabling Efficiency, Speed & Accuracy

Intelligent Analysis, Decision & Execution

We build AI systems for teams that need search across company data, workflow automation across tools, or AI features inside products and internal software.

Talk to Us

Search across company data

Give teams answers from docs, tickets, runbooks, and product data with sources and permissions.

Useful when people spend too long searching or get different answers from different systems.

Enterprise searchRAGPermissions

Automate internal workflows

Use AI to route work, draft outputs, trigger actions, and keep approvals and logs in place.

Useful when repetitive work moves across multiple tools and teams.

AI agentsWorkflow automationGovernance

Add AI to products and internal tools

Build assistants, guided actions, or decision support into the software your team or customers already use.

Useful when AI needs to be part of the product, not a separate tool.

AI integrationDecision supportModel routing

ODIN DETECTOR ECOSYSTEM

Related Terms

ODIN (Out-of-DIstribution detector for Neural networks) leverages temperature scaling and adversarial perturbation to separate known from unknown signals. These related concepts form the mathematical and architectural foundation for its operation.

Temperature Scaling

A confidence calibration technique that divides the output logits by a learned scalar parameter T. When T > 1, the SoftMax distribution softens, amplifying the separation between the maximum probability of in-distribution and out-of-distribution samples. ODIN uses this as a preprocessing step before applying perturbations, making the model's overconfidence on unknowns collapse more dramatically than on known classes.

T > 1

Optimal Temperature Range

Adversarial Perturbation

ODIN adds a small, targeted FGSM-style perturbation to the input before classification. The perturbation is calculated as the sign of the gradient of the SoftMax score with respect to the input, pushing the sample slightly in the direction that increases the score of its predicted class. In-distribution samples resist this perturbation, while out-of-distribution samples exhibit a much larger change in their SoftMax output, creating a clear separability gap.

Perturbation Magnitude

SoftMax Score Thresholding

The final detection mechanism in ODIN. After temperature scaling and adversarial perturbation, the maximum SoftMax probability is computed. A simple threshold is applied: if the calibrated and perturbed maximum score falls below the threshold, the sample is flagged as out-of-distribution. This non-parametric decision rule requires no retraining and can be tuned to balance the true positive rate against the false positive rate for unknown modulation rejection.

No Retraining

Deployment Cost

Outlier Exposure

A complementary training-time regularization technique that improves ODIN's baseline performance. By exposing the model to a diverse auxiliary dataset of outlier examples during training, the network learns a tighter decision boundary around known classes. When combined with ODIN's post-hoc detection, models trained with outlier exposure achieve state-of-the-art AUROC and FPR at 95% TPR metrics on standard OOD detection benchmarks.

EXPLORE

Mahalanobis Distance

An alternative scoring function to the SoftMax threshold used in ODIN. Instead of relying on the maximum class probability, this method fits a class-conditional Gaussian distribution to the features of each known class at a specific layer. The Mahalanobis distance from a test sample to the nearest class centroid provides a more statistically robust OOD score. ODIN's perturbation and temperature scaling can be applied as preprocessing to enhance this distance-based detection as well.

Feature Space

Detection Layer

Energy-Based Detection

Uses the Helmholtz free energy function E(x) = -T * log(sum(exp(f_i(x)/T))) as an OOD score, where f_i(x) are the logits. This energy score is theoretically aligned with the input's probability density and is less susceptible to feature collapse than SoftMax. ODIN's temperature parameter T directly modulates this energy function, and the adversarial perturbation step can be reinterpreted as moving samples along the energy landscape to widen the gap between in-distribution and OOD energy scores.

About the author

Prasad Kumkar

CEO & MD, Inference Systems

Prasad Kumkar is the CEO & MD of Inference Systems and writes about AI systems architecture, LLM infrastructure, model serving, evaluation, and production deployment. Over 5+ years, he has worked across computer vision models, L5 autonomous vehicle systems, and LLM research, with a focus on taking complex AI ideas into real-world engineering systems.

His work and writing cover AI systems, large language models, AI agents, multimodal systems, autonomous systems, inference optimization, RAG, evaluation, and production AI engineering.

Limited slotsGet a Free AI Consultation

How We Work

Custom AI workflows for your Business

One-fit-all AI don't work for modern businesses. At Inferensys, we aim to understand your business & custom requirements; which we use to define most efficient agentic workflows, the data, and the tools for your business.

Talk to Us

ODIN

What is ODIN?

Key Features of ODIN

Temperature Scaling

Input Preprocessing with Adversarial Perturbations

Threshold-Based Detection

Hyperparameter Tuning Protocol

Compatibility with Pre-Trained Models

Performance Benchmarks

Frequently Asked Questions

Intelligent Analysis, Decision & Execution

Search across company data

Automate internal workflows

Add AI to products and internal tools

Outlier Exposure

Prasad Kumkar

Partnered with leading AI, data, and software stack.

Custom AI workflows for your Business

Review the use case

Pick the right approach

Build the first useful version

Improve from there