MONTE CARLO DROPOUT BASED ACTIVE LEARNING FOR DEEP LEARNING IN STRUCTURAL SIMULATION

In recent years, deep learning (DL) has been widely used in the realm of maritime technology and engineering, particularly in the domain of structural simulation. However, one of the significant challenges in this context is the requirement for extensive training data, which is cost-prohibitive. To mitigate the expensive cost of generating training data, an active learning (AL) model is proposed to reduce the required number of labeled samples for deep learning. Within this active learning model, a predefined acquisition function is used to query an unlabeled sample pool. It focuses on selecting highly informative samples for labeling and subsequent inclusion into the training dataset. The predefined oracle in the proposed active learning model is developed based on the Monte Carlo dropout technique, which is used to quantify the aleatory uncertainty and epistemic uncertainty inherent in the deep learning model. Consequently, samples characterized by maximal uncertainty are identified and subsequently integrated into the training dataset. A comparative study with random sampling is conducted to demonstrate the effectiveness and advantage of the proposed active learning model. This model can be generalized to improve other deep learning models used in the maritime technology and engineering domain, as it offers a computationally efficient way of sampling additional data.