Operationalizing deployed machine learning models in safety-critical or other applications through out-of-distribution robustness quantification

By generating an OOD score from intermediate transformer layer feature outputs and controlling output passage based on this score, the method addresses illogical responses in transformer-based models, enhancing their robustness and reliability in safety-critical applications.

WO2026151660A1PCT designated stage Publication Date: 2026-07-16RAYTHEON CO
View PDF 0 Cites 0 Cited by

Patent Information

Authority / Receiving Office
WO · WO
Patent Type
Applications
Current Assignee / Owner
RAYTHEON CO
Filing Date
2026-01-05
Publication Date
2026-07-16

AI Technical Summary

Technical Problem

Transformer-based machine learning models, such as large language models, often produce illogical, incorrect, or misleading responses when operating out-of-distribution, posing challenges in safety-critical applications due to unaccounted input variations during training.

Method used

Implementing a method to generate an out-of-distribution (OOD) score by analyzing feature outputs from intermediate transformer layers, comparing it against a threshold, and controlling the passage of outputs to downstream processes based on this score, using techniques like network confidence, feature distance, or likelihood-based approaches.

Benefits of technology

Enhances the robustness of transformer-based systems by accurately identifying in-distribution and out-of-distribution inputs, ensuring reliable and trustworthy outputs in safety-critical applications.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure US2026010169_16072026_PF_FP_ABST
    Figure US2026010169_16072026_PF_FP_ABST
Patent Text Reader

Abstract

A method (300) includes receiving (305) an input (151) for a machine learning model (110) having a transformer-based architecture with an embedding layer (111), an unembedding layer (113), and three or more transformer layers (112). The method also includes embedding (310) the input at the embedding layer and passing (310) the embedded input from the embedding layer across the transformer layers. Each transformer layer provides one or more feature outputs (114a, 114b). The method further includes obtaining (315) the feature output(s) from a first intermediate transformer layer. The method also includes generating (320) an out-of-distribution (OOD) score (155) for the received input based on the feature output(s) from the first intermediate transformer layer and one or more feature distributions (125a, 125b) from the first intermediate transformer layer. In addition, the method includes comparing (325) the OOD score against a threshold and, in response to determining that the OOD score exceeds the threshold, passing (330) one or more outputs from the unembedding layer to one or more downstream processes (160).
Need to check novelty before this filing date? Find Prior Art