Abstract:

Artificial Neural Networks (ANNs) have achieved extraordinary success in data processing applications ranging from image recognition to time series prediction. The availability of vast datasets for training, as well as the increasing complexity of the models, may be credited to the success. Unfortunately, only a limited number of examples are provided for training in certain applications. In high-complexity models, fewer training samples increase the risk of over-fitting and poor generalisation. Furthermore, as compared to simpler models, complicated models with a high number of trainable parameters take more energy to train and optimise. To the best of our knowledge, this study proposes the first application of ANNs for Early Stage Alzheimer Disease (ES-AD) classification from handwriting (HW). We suggest utilising Reservoir Computing (RC), a methodology for creating Recurrent Neural Networks (RNNs) that simplifies training by optimising just the output layer, both numerically and empirically. For comparison, we also present the Bidirectional Long Term Short Term (BiLSTM) and Convolutional Neural Network (CNN) approaches. In order to examine the accuracyefficiency trade-off, we consider not only the accuracies but also the energy expenses necessary to acquire the various accuracies. Our numerical and experimental findings reveal that RC achieves a classification accuracy of 85%, which is 3% lower than BiLSTM and 2% higher than CNN, while requiring substantially less training and much less inference. We expect that our results emphasise the need of investigating the accuracy-efficiency tradeoff of different models in the community in order to lessen the overall environmental effect of ANNs training.

Description:

Artificial neural network, Early stage alzgeimer disease classification, recurrent neural network, reservoir computing