A Novel One-Versus-All Approach for Multiclass Classification in TinyML Systems

The recent progress in TinyML technologies triggers the need to address the challenge of balancing inference time and recognition quality. TinyML systems are defined by specific constraints in computation, memory and energy. These constraints emphasize the need for specialized optimization techniques when implementing machine learning (ML) applications on such platforms. While deep neural networks are popular in TinyML systems, exploring simple classifiers is also worthwhile. In this work, we consider a modification of the one-versus-all (OVA) approach in a multiclass task of computer vision in TinyML systems. This modification, named thresholded OVA (TOVA), enables control over classification accuracy, influencing both latency and energy consumption per inference. By testing various combinations of hyperparameters, we simulate the performance of a real device using metrics specific to TinyML systems. The results show that the proposed method significantly saves energy and speeds up computation, at the cost of slightly lower-overall accuracy of the TinyML system.