Uncertainty Analysis for Deep Learning Prediction in Human Activity Recognition based on Monte Carlo Approach

This paper explores uncertainty estimation in neural network predictions for human activity recognition using a Monte Carlo approach. An accelerometer embedded in a wearable device on the user's chest provides input features such as acceleration along three axes for the deep learning model. Monte Carlo Dropout is employed during both training and inference to simulate multiple forward passes, enabling different predictions for the same input. This method allows for detailed analysis and visualization of prediction uncertainty. Among all activities, cough and fall are two main classes which are under significant concern, so the variation of each input feature will be examined by injecting noise variation to each acceleration on X, Y and Z axis subsequently. New noisy data will be generated based on the data from a real dataset, then Monte Carlo Dropout model provides the predicted probability of the cough and fall class along with their associated 95% confidence interval for uncertainty based on each type of acceleration fluctuation. The average width of the confidence intervals of all new generated samples are calculated to identify the feature with the lowest uncertainty, which would generally deliver the most reliable predictions for the given class. The utilized approach is highly essential when the obtained dataset has limited amount of data.