Confidence score - True Class Probability (TCP)

True Class Probability (TCP)[Corbiere et al., 2021] is a confidence score method that aims to flag whether a neural network prediction is conform to the true class probability distribution. It is a module that can be plugged on a neural network, and only requires the output of a given layer for training and operational phase. Specifically, TCP requires access to
the penultimate layer, which outputs the probability distributions (the logits).Training TCP boils down to a classification problem, which requires to have a labelled dataset (although it only requires the original labels; no additional labelling effort is required).

It is dedicated to classification and detection use cases.

Documentation

Methodological Guidelines

• Methodological guideline for Robustness Functional Set
• Methodological Guideline and Evaluation tools for robust artificial intelligence

Benchmark

• Report on Confidence Score improvement methods

State of the art

• Characterization of the notion of trust
• State of the Art on Data based AI Confidence Evaluation

Support

Support for True Class Probability must be obtained by sending an email to support@confiance.ai

Ensure your email contains :

• Your name
• A link to this page
• the version you are working with
• A clear description of the problematic (bug, crash, feature or help request)
• A full description of the problem whichallow to reproduce it
• Any file or screenshort element mandatory for the full understanding of the problematic