A Fast Forward Algorithm to Learn the Classification of Objects In Space

The forward-forward (FF) algorithm has recently emerged as a promising alternative to backpropagation for training deep neural networks, performing two forward passes using real (positive) data and synthetic (negative) data. Building on this paradigm, we introduce "A Fast Forward Algorithm to Learn the Classification of Objects iN Space (AF: FALCONS)", which is a hybrid learning framework that integrates decision tree structures into the FF training process for efficient and accurate space object classification. Our method enhances the learning and classifying space objects from positive samples while mitigating the influence of low-goodness negative samples. Thus, we demonstrate that the proposed AF: FALCONS can enhance the classification of space objects, thereby showing a promising avenue for space situational awareness tasks. We conduct numerical simulations based on the proposed learning algorithm on the SPARK (SPAcecraft Recognition leveraging Knowledge) dataset and benchmark it against both standalone FF networks and conventional backpropagation-trained feedforward neural networks (FFNNs). Numerical simulations have shown that the proposed learning algorithm achieves reduced training error and higher testing accuracy, while highlighting its effectiveness and computational efficiency in training and learning. With these computational efficiencies, the proposed AF: FALCONS architecture is particularly well-suited for deployment on spacecraft with limited processing power, enabling onboard object recognition and classification purposes, which are in need of critical autonomous space operations.