Monte Carlo Concrete DropPath for epistemic uncertainty estimation in pollen images classification

The paper presents the results of a new method for training the NASNet neural network called Monte Carlo Concrete DropPath for epistemic uncertainty estimation to classify pollen images. The developed method is compared with existing methods for epistemic uncertainty estimation. The method turns an arbitrary multipath neural network into a Bayesian one by sampling from the predictive distribution. The Monte Carlo method samples different masks of DropPath to estimate uncertainty. Moreover, the probability of DropPath is optimized using continuous relaxation. The proposed method was tested for the classification task on the state-of-the-art NASNet architecture. The method demonstrated advantages on the task of classifying pollen images. The classification accuracy increased for 13 pollen species of allergen plants by 0.73 % on average compared to the baseline NASNet, reaching 98.34 % by F1 measure. Furthermore, the method increased calibration and reduced the epistemic uncertainty of the model by two times compared to the NASNets ensemble. It is shown that continuous relaxation of the DropPath probability parameter increases the accuracy of problem solving and reduces the epistemic uncertainty of the model. These results contribute to the automation of aeropalinological monitoring to reduce the time of informing patients who suffer from pollinosis and hence to prevent allergy symptoms. The developed method can be applied to train a neural network for other computer vision tasks on any image dataset.