Toward Multi-Level Automatic Updates of Deep Learning Models Using New Micro, Meso, and Macro Stability Metrics

Training deep learning models is a reasonably complex task. This complexity may drastically increase, particularly within the context of complicated industrial ecosystems. However, far beyond this complexity, major challenges continue to concern the maintenance of deep learning models. More precisely, it is extremely complex to maintain a model reliable throughout its life cycle. Indeed, frequent updates may not necessarily result into better performance, compared to older versions of the model. Therefore, it is of paramount importance to possess a solution that can suggest the right updates and the right time in a given deep learning model. To meet this goal, we propose in this paper a new approach for an automatic updating of deep learning models. Our approach includes eight new metrics that assess the stability of the models in terms of the disparity between predictions and/or versions. The metrics will ultimately allow for multi-level updates of the learning models. The right updates will, therefore, be implemented at the right time. In order to demonstrate the performance of our solution, we conducted several experiments on Convolutional Neural network architectures using the MedM-NIST dataset. Our current simulations are showing promising results in the specific field of predicting survival from colorectal cancer histology slides.