Device and Computer Implemented Method for Machine Learning

Simple SummaryContent extracted from patent full text and abstract with AI.

This patent describes both a device and a computer-implemented method for training artificial neural networks to be more robust in real-world scenarios. The key idea is to train the neural network not just on normal data, but also considering the largest possible 'perturbations' (small changes or noise) in the input data that still lead to correct outputs. This approach helps prevent drops in accuracy for data unseen during training, making machine learning systems more reliable under varied and potentially challenging real-world conditions.

Use CasesContent extracted from patent full text and abstract with AI.

• Autonomous vehicles, such as self-driving cars, for detecting traffic signs, pedestrians, and road conditions.
• Robotics, where robots must recognize and adapt to objects and environments with uncertainty or sensor noise.
• Industrial automation, improving machine vision for manufacturing defect detection despite variability in input data.
• Smart home appliances that use computer vision or audio recognition for operation in unpredictable environments.
• Access control systems relying on face, voice, or object recognition for secure entry.
• Healthcare devices using image or signal classification in unpredictable, non-ideal conditions.

BenefitsContent extracted from patent full text and abstract with AI.

• Increases robustness of machine learning models against unseen or adversarial data.
• Minimizes performance drops caused by input variations or sensor noise.
• Improves reliability and safety of AI-powered systems in critical real-world applications.
• Adaptable across sensor modalities (vision, radar, audio, thermal, etc.) and application domains.
• Reduces risk of misclassification in safety-critical scenarios (e.g., autonomous driving).
• Provides a systematic way to balance accuracy and robustness during neural network training.

Technical Classifications (CPCs)

Inventors & Applicants

Patent Abstract

A device (100) and a computer implemented method for machine learning, wherein the method comprises providing an artificial neural network (106) that is configured to determine an output of the artificial neural network (106) depending on an input of the artificial neural network (106) and depending on weights of the artificial neural network (106), providing a sample for the input and a ground truth for the output, wherein the input is defined depending on the sample and depending on a perturbation, determining a perturbation, in particular a largest perturbation, for that the output indicates the ground truth, determining the weights of the artificial neural network (106) depending on a loss that is defined depending on the output and the ground truth.