Beyond-Backpropagation Training: Methods, Applications, and Perspectives

Artificial intelligence has revolutionized numerous domains, including computer vision, natural language processing, and autonomous systems. These breakthroughs have been largely driven by the backpropagation (BP) algorithm, which allows neural networks to be trained efficiently. However, BP suffers from limited biological plausibility, computational inefficiency, and memory bottleneck, all of which hinder its applicability to neuromorphic, edge, and large-scale distributed systems. To address these limitations, numerous BP-free algorithms have emerged over the past decade. The rapid growth and variety of these approaches call for a systematic review of existing BP-free methods. In response, this work provides a comprehensive survey of BP-free learning methods, covering representative algorithms across diverse domains. Specifically, we propose a unified taxonomy based on two dimensions-how update signals are computed, and when updates are applied. Building on this taxonomy, we analyze fundamental properties of BP-free methods and compare their performance against BP counterparts, which offers practical guidelines for selecting these methods across different domains. We further highlight key open challenges for future research, including improving theoretical understanding, scaling to deeper networks, and developing biologically inspired learning rules built upon brain-like architectures. We hope this work could provide a roadmap for researchers exploring biologically plausible and resource-efficient training.