Resource-Constrained Implementation of Deep Learning Algorithms for Dynamic Touch Modality Classification

Integrating Machine Learning (ML) algorithms with tactile sensing arrays yield sophisticated systems capable of performing intelligent tasks. Such systems can be used in prosthetic devices and robotics applications, enabling conducting daily tasks and manipulations. This paper presents low-cost and resource-constrained implementations of deep learning algorithms for the classification of dynamic touch modality based on alphabetic letter patterns. This work provides a comparison between two types of deep neural networks: 1-D Convolutional Neural Networks (CNNs) and Recurrent Neural Networks (RNNs). Moreover, the models providing the best performance in terms of accuracy and computational cost have been deployed on a resource-constrained embedded system. Experimental results show that 1-D CNN outperforms RNNs in terms of both accuracy and computational cost achieving a classification time of 242 ms using 32-bit floating point on the Arduino Nano BLE hardware device.