Method and system to classify sensor data with improved training robustness

A classifying system configured to classify sensor data with improved robustness against label noise, the system comprising:

• a training data storage device configured to store training data, the training data including multiple training input data and corresponding labels, the training input data representing physical properties of a physical system obtained from one or more sensors;
• a base classifier defined for the training data, the base classifier including an encoder and a classifier, the encoder being configured to map an input data to a latent representation, the classifier being configured to be applied to the latent representation wherein the encoder is defined by parameters that are pre-trained, a base training function is configured for optimizing parameters defining the classifier according to the training data;
• a novel input data interface configured to obtain a novel input data;
• and
• a processor system configured to:
• define a noise level for the training labels,
• compute a predicted label for the novel input data with improved robustness against label noise the computing including estimating a label which is most likely under repeated application of the base training function to the training labels incorporating noise according to the noise level and subsequent application of the base classifier configured, according to a base prediction function, to the novel input
• compute a radius indicating a number of label changes for which the predicted label for the novel input will not change if a number of label changes does not exceed the indicated number of label changes,
• obtain an estimate of a number of corrupted labels in the training data, and
• compare the estimated number of corrupted labels with the radius before accepting the predicted label for the novel input.

A non-transitory computer readable medium on which is stored data representing instructions for classifying sensor data with improved robustness against label noise, the instructions, when executed by a computer system, causing the computer system to perform the following steps:

• providing training data, the training data including multiple training input data and corresponding training labels, the training input data representing physical properties of a physical system obtained from one or more sensors;
• defining a base classifier for the training data, the base classifier including an encoder and a classifier, the encoder being configured to map an input data to a latent representation, the classifier being configured to be applied to the latent representation wherein the encoder is defined by parameters that are pre-trained, a base training function is configured for optimizing parameters defining the classifier according to the training data;
• providing a novel input data;
• defining a noise level for the training labels;
• computing a predicted label for the novel input with improved robustness against label noise by estimating a label which is most likely under repeated application of the base training function to the training labels incorporating noise according to the noise level and subsequent application of the base classifier, configured according to a base prediction function, to the novel input;
• computing a radius indicating a number of label changes for which the predicted label for the novel input will not change if a number of label changes does not exceed the indicated number of label changes;
• obtaining an estimate of a number of corrupted labels in the training data;
• and
• comparing the estimated number of corrupted labels with the radius before accepting the predicted label for the novel input.