Exploring the Forward-Forward Algorithm to Train Neural Networks for Camera Trap Images on the Edge

Camera trap systems and motion-activated field cameras are edge devices commonly used for animal monitoring in the wild. Using such technologies has revolutionized ecological research and wildlife monitoring, enabling non-invasive study of animal movement, behavior, and population dynamics. However, the manual analysis and classification of the vast number of camera trap images can be time-consuming and challenging for researchers. The image data from these cameras is also naturally class unbalanced because some species are rarer than others, and typically more than half the images are “ghost” images triggered by anomalous events like the wind. Neural networks using the backpropagation (BP) algorithm have been used to classify camera trap images. Since the deployment context of each camera trap is unique, there is a need to custom-train the classification model for each camera on the edge. Unlike BP, the newly proposed Forward-Forward (FF) algorithm reduces the memory needs for training on the edge. This paper explores how well the FF algorithms could distinguish between “ghost” and “animal” images for camera trap images when trained on the edge. FF algorithm on different image sizes and varying the number of neurons in each layer is investigated. The primary conclusion is that both BP and FF have similar performance and inference time on this task. In addition, FF is slower to train but requires 60% less memory than BP while training.