Towards a Trade-off Between Accuracy and Computational Cost for Embedded Systems: A Tactile Sensing System for Object Classification

The deployment of the inference phase in self–standing systems, which have resource–constrained embedded units, is faced with many challenges considering computational cost of the elaboration unit. Therefore, we propose using a learning strategy based on a loss function that leads to finding the best configuration of the prediction model balancing the generalization performance and the computational cost of the whole elaboration system. We validate our proposal by integrating a tactile sensing system on a Baxter robot to collect and classify data from five daily–life objects using four different algorithms. Results show that the best performance, when the computational cost is not relevant, is achieved by the fully–connected neural network using 16 features, while, when the computational cost matters, the loss function showed that the kernel SVM with 4 features has the best performance.