Sparse Forward-Forward Algorithm: Efficient Machine Learning for Resource-Constrained Edge Devices

Deploying machine learning models on resource-constrained edge devices, such as microcontroller, is central to EdgeML. However, traditional training algorithms like backpropagation are computationally intensive and energy demanding. We propose the Sparse Forward-Forward (SFF) Algorithm, a novel training approach that reduces computational effort by selectively processing only the top-$k$ most significant activations during training, maintaining accuracy while improving efficiency. We evaluate SFF on five benchmark datasets—MNIST, Fashion MNIST, Human Activity Recognition (HAR), Speech Commands, and Flower—against backpropagation (BP) and the standard Forward-Forward (FF) algorithm. Results show that SFF achieves comparable accuracy with substantial efficiency gains: up to $20 \times$ lower effort than BP on MNIST and up to $61 \times$ on the Flower dataset, demonstrating its suitability for energy-constrained edge scenarios. SFF offers a trade-off between performance and resource usage. It is well-suited for scalable EdgeML and Industrial Cyber-Physical Systems (ICPS), where local adaptability and low-latency learning are essential.