Tiny Neural Deep Clustering: An Unsupervised Approach for Continual Machine Learning on the Edge

Traditional tiny machine learning systems are widely employed because of their limited energy consumption, fast execution, and easy deployment. However, such systems have limited access to labelled data and need periodic maintenance due to the evolution of data distribution (i.e., context drift). Continual machine learning algorithms can enable continuous learning on embedded systems by updating their parameters, addressing context drift, and allowing neural networks to learn new categories over time. However, the availability of labelled data is scarce, limiting such algorithms in supervised settings. This paper overcomes this limitation with an alternative approach which combines supervised deep learning with unsupervised clustering to enable unsupervised continual machine learning on the edge. Tiny Neural Deep Clustering (TinyNDC) is deployed in an OpenMV Cam H7 Plus and tested with the MNIST dataset reaching a classification accuracy of 92.3% and a frame rate of 44 FPS.