Incremental learning for a flexible CAD system design
Deep neural networks suffer from Catastrophic Forgetting (CF) on old tasks when they are trained to learn new tasks sequentially, since the parameters of the model will change to optimize on the new class. The problem of alleviating CF is of interest to Computer aided diagnostic (CAD) systems community to facilitate class incremental learning (IL): learn new classes as and when new data/annotations are made available and old data is no longer accessible. However, IL has not been explored much in CAD development.
We propose a novel approach that ensures that a model remembers the causal factor behind the decisions on the old classes, while incrementally learning new classes. We introduce a common auxiliary task during the course of incremental training, whose hidden representations are shared across all the classification heads. Since the hidden representation is no longer task-specific, it leads to a significant reduction in CF. We demonstrate our approach by incrementally learning 5 different tasks on Chest-Xrays and compare the results with the state-of-the-art regularization methods. Our approach performs consistently well in reducing CF in all the tasks with almost zero CF in most of the cases unlike standard regularisation-based approaches.
Manifold learning to address catastrophic forgetting
A major challenge that deep learning systems face is the Catastrophic Forgetting (CF) phenomenon that is observed when fine-tuning is used to try and adapt a system to a new task or a sequence of datasets with different distributions. CF refers to the significant degradation in performance on the old task/dataset. In this paper, a novel approach is proposed to address CF in computer aided diagnosis (CAD) system design in the medical domain. CAD systems often need to handle a sequence of datasets collected over time from different sites with different imaging parameters/populations.
The solution we propose is to move samples from all the datasets closer to a common manifold via a reformer at the front end of a CAD system. The utility of this approach is demonstrated on two common tasks, namely segmentation and classification, using publicly available datasets. Results of extensive experiments show that manifold learning can yield about 74% improvement on an average in the reduction of CF over the baseline fine-tuning process and the state-of-the-art regularization based methods. The results also indicate that a Reformer when used in conjunction with the state-of-the-art regularization methods, has the potential to yield further improvement in CF reduction.