Terminal phase of training Decision Support Tool (TDST)

AFRL/RD has an interest in the phenomenon of double descent or “neural collapse” (NC) in which a deep learning model’s training loss drops dramatically after an extended period of flattening. If a model training is ended before double decent occurs, as is typical with early stopping, the global minimum will not be reached by the optimization algorithm and a suboptimal solution will result, leading to a model will fit to make accurate inferences (e.g., object classification or segmentation). By extension, since the Air Force requires reliable and accurate deep learning models for directed energy (DE) applications, explaining and predicting neural collapse is important. Spectral Sciences, Inc. (SSI), in collaboration with DE experts at the Air Force Institute of Technology (AFIT), propose to develop a software tool called Terminal phase of training Decision Support Tool” (TDST). TDST works by monitoring multiple layers of the network during training, evaluating the likelihood double decent will occur and providing the model engineer quantitative metrics to consider. In the process, we also plan to interpret how and why neural collapse occurs, contributing to existing research in explainable AI..  Phase I will result in a demonstration showing double descent and quantifying its effect on DE-relevant datasets, and the development of a prototype explainability analysis tool. In Phase II, the software tool will be expanded to include the prediction of double descents that may be possible during training but have not yet been reached. Additionally, the tool will be expanded to characterize and quantify the specific type of collapse expected. The software will be generalized for deployment across different neural network architectures and datasets. Phase III will expand the Technology Readiness Level (TRL) of TDST into a full end-to-end DE simulation framework.