Towards the on-device Handwriting Trajectory Reconstruction of the Sensor Enhanced Pen

Performing handwriting trajectory regression from inertial data using Deep Neural Network (DNN) on an embedded device is a very challenging task, since the network accuracy is prone to imperfections in the weights and needs a significant amount of parameters to be able to regress. In this work, we apply and compare different quantization techniques and Mitchell logarithmic multiplication approximation in order to enable the on-device inference. We show that it is possible to perform the inference of the TCN-based regression model using only 8-bit fixed-point quantization without significant reconstruction precision loss and that the accuracy degradation of the approximate multiplication can be partially compensated with Quantization-aware Training (QAT). Finally, we demonstrate that the compressed models can be integrated into an off-the-shelf commercial Systems-on-Chip with minimal use of FPU and requiring only 460 KB of the ROM size for the TCN-49 configuration.