Brief

Researchers have introduced a new framework called Uncertainty Activation Map (UAM) to visualize uncertainty in deep learning models. This method combines Evidential Deep Learning with Full-Gradient Class Activation Mapping to create spatial maps that highlight areas of missing or conflicting evidence within input data. The UAM framework distinguishes between vacuity (lack of evidence) and dissonance (conflicting evidence), offering a more interpretable understanding of model reliability for safety-critical applications. AI

A new framework called PaTAS has been developed to model and propagate trust within neural networks using Subjective Logic. This system operates in parallel with standard neural computations, employing Trust Nodes and Trust Functions to assess and transmit trust levels related to inputs, parameters, and activations. PaTAS includes mechanisms for refining parameter reliability during training and for calculating instance-specific trust during inference, demonstrating its ability to provide interpretable trust estimates that complement accuracy metrics and highlight reliability issues in data. AI

Researchers have developed a new method using Subjective Logic to dynamically update confidence in AI safety arguments during runtime. This approach integrates evidence from both the design phase and real-time performance indicators to continuously assess and adjust safety claims. The system is designed to be responsive, penalizing violations promptly while increasing confidence when safety is maintained, as demonstrated with a simulated construction zone assist function. AI