Method of training a machine learning model

EP4728432A1Pending Publication Date: 2026-04-22NEC ONCOIMMUNITY AS
0 Cites 0 Cited by

Patent Information

Authority / Receiving Office
EP · EP
Patent Type
Applications
Current Assignee / Owner
NEC ONCOIMMUNITY AS
Filing Date
2024-06-13
Publication Date
2026-04-22

AI Technical Summary

Technical Problem

Machine learning models face challenges in training when verified negative data is limited or unavailable, particularly in fields like B-Cell epitope prediction, where confirming the absence of a characteristic is impractical at scale.

Method used

A method involving the generation of pseudo-random data for training, where two epochs of machine learning training use pseudo-random data as negative training data, allowing the model to robustly learn without relying on confirmed negative examples, thereby reducing false negatives and skewing.

Benefits of technology

This approach enables robust training of machine learning models even in the absence of verified negative data, improving model performance and reducing the impact of false negatives, leading to more accurate predictions in fields like B-Cell epitope identification.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure EP2024066409_19122024_PF_FP_ABST
    Figure EP2024066409_19122024_PF_FP_ABST
Patent Text Reader

Abstract

A method of training a machine learning model is disclosed, comprising: generating (204) first pseudo-random data (116); performing (206) a first pass of machine learning training, using positive training data (114) and using the first pseudo-random data as negative training data, to complete a first epoch; generating (208) second pseudo-random data; and performing (210) a second pass of machine learning training, using the second pseudo-random data as negative training data and using the positive training data, to complete a second epoch; wherein generating both the first pseudo-random data and the second pseudo-random data comprises generating pseudo-random data in different ways, including generating completely pseudo-random data and generating pseudo-random data by modifying each sample of the positive training data.
Need to check novelty before this filing date? Find Prior Art