Machine Learning Hardware Implementation of Handwritten Digit Inference using Arduino and Ternary Output Binary Neural Network

In this paper, we analyze binary neural network (BNN) and ternary output BNN (ToBNN) from a software perspective, and introduce tiny machine learning (TinyML) hardware implementation of handwritten digit inference. Both BNN and ToBNN achieve a reduction of approximately 70% in memory usage for weight storage by using binary values. Inference energy consumption was also reduced by approximately 32%. ToBNN achieves 90.24% inference accuracy for MNIST dataset, while conventional BNN shows 50.25% accuracy. To implement TinyML hardware, Arduino Uno, Portenta H7, TFT touch LCD shield, and 16×2 LCD were used. The hardware implementation results show successful inference for handwritten digits.