Likelihood Ratio OOD Detection for Background-Shifted Inputs

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing deep generative model-based approaches for out-of-distribution (OOD) detection fail to distinguish between semantic content relevant for classification and background statistics, often assigning higher likelihood to OOD inputs, leading to erroneous classifications.

Innovation Solution

A likelihood ratio method using a semantic model trained on in-distribution data and a background model trained on perturbed in-distribution data to correct for background statistics, enhancing OOD detection by calculating a likelihood ratio that isolates semantic components.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Reliability

If a deep generative model is trained on training data to detect OOD inputs, then the model can identify anomalous data, but the model assigns higher likelihood to OOD inputs than in-distribution inputs, leading to erroneous classifications

Engineering Contradiction:
ImproveOOD detection accuracyVSAvoidlikelihood assignment accuracy
Core Design Contradiction:
ReliabilityVSMeasurement precision

Solution Approach 1:

The patent segments the likelihood evaluation into two separate components: a semantic model that evaluates class-specific features and a background model that evaluates overall data distribution. By dividing the single generative model into these two specialized components, the system can separately assess semantic relevance and distributional likelihood, preventing OOD inputs from receiving erroneously high likelihood scores while maintaining accurate in-distribution classification.

Inventive Principle:
Principle #1Segmentation

2Measurement precision

If a neural network classifier is deployed to classify bacterial genomes, then high accuracy is achieved on training classes, but the classifier wrongly classifies OOD inputs from unseen bacterial classes with high confidence

Engineering Contradiction:
Improveclassification accuracyVSAvoidhandling of unseen classes
Core Design Contradiction:
Measurement precisionVSAdaptability or versatility

Solution Approach 1:

The patent introduces an intermediary OOD detection layer between the neural network classifier and the final classification output. This intermediary component evaluates whether an input belongs to any known in-distribution class before allowing classification. By adding this intermediate check, the system maintains high classification accuracy for known classes while preventing confident misclassification of unseen bacterial classes as known classes.

Inventive Principle:
Principle #24Intermediary (Mediator)

3Productivity

If the model processes all inputs through full classification, then comprehensive analysis is performed, but computing resources are wasted on OOD inputs that cannot be correctly classified

Engineering Contradiction:
Improveprocessing throughputVSAvoidcomputing resource consumption
Core Design Contradiction:
ProductivityVSLoss of energy

Solution Approach 1:

The patent performs preliminary OOD detection before executing the full classification pipeline. By first evaluating whether an input belongs to the training data distribution using the background model, the system can early-screen and exclude OOD inputs from resource-intensive full classification processing. This preliminary action significantly reduces computing resource consumption on OOD inputs while maintaining comprehensive analysis for valid in-distribution inputs.

Inventive Principle:
Principle #10Preliminary action

Data Source

PatentEP3980941B1Likelihood ratios for out-of-distribution detection
Publication Date: 2025.11.19 GOOGLE LLC
  • EP3980941B1 patent drawingFigure 1A
  • EP3980941B1 patent drawingFigure 1B
  • EP3980941B1 patent drawingFigure 1C

AI summary

The present disclosure is directed to systems and method to perform improved detection of out-of-distribution (OOD) inputs. In particular, current deep generative model-based approaches for OOD detection are significantly negatively affected by and struggle to distinguish population level background statistics from semantic content relevant to the in-distribution examples. In fact, such approaches have even been experimentally observed to assign higher likelihood to OOD inputs, which is opposite to the desired behavior. To resolve this problem, the present disclosure proposes a likelihood ratio method for deep generative models which effectively corrects for these confounding background statistics.