Deep learning is a class of machine learning algorithms that utilizes multiple layers to extract higher-level features from the raw input progressively. Deep learning is extremely beneficial in collecting, handling, interpreting, and analyzing a vast amount of data efficiently. Deep learning owns the ability to extract features automatically and is powerful in handling a huge volume of unstructured data. Uncertainty is a crucial problem in data handling and caused due to unknown or imperfect data sets. In deep learning, uncertainty arises when there is no appropriate training data or testing, and training data are mismatched.
Uncertainties occur in numerous real-world applications such as finance, marketing, medical diagnosis, weather forecasting, to name a few. Such sudden changes in the deep learning models affect performance and accuracy. It is necessary to predict and handle uncertainties, and one of the prediction classification models to detect the uncertainty is evidential deep learning. Evidential deep learning observes and measures the uncertainty in datasets with evidence theory.
• Existing deep neural networks fail to estimate the predictive uncertainty for a classification problem due to their utilization of the softmax layer, and it tends to be over-confident in false prediction.
• To address this problem, Evidential Deep Learning (EDL) is a recent approach developed to overcome the limitations of softmax-based DNNs by training the single neural networks to estimate predictive uncertainties based on Dempster-Shafer Theory (DST).
• EDL provides a principled way to formulate the multi-class classification and uncertainty modeling jointly.
• Evidential deep learning extends the idea of learning the probability distribution parameters to predict higher-order distributions over the original likelihood parameters themselves.
• Deep evidential learning improves its performance as efficient and reliable with better accuracy through the Multitask learning framework.
• In Open Set Action Recognition (OSAR), the potential over-fitting issue hampers the generalization capability for achieving good OSAR performance because of the deterministic nature of EDL.
