Is Backpropagation Optimal? When Synthetic Gradients Improve Sample Efficiency

A new theoretical analysis challenges the long-held dominance of backpropagation in artificial neural networks, suggesting that synthetic gradients can offer superior sample efficiency under specific conditions, according to a paper submitted on 27 May 2026 [1][2]. Backpropagation has been the standard learning rule for training artificial neural networks, a class of computational models inspired by biological brains that consist of layers of interconnected nodes [1][4]. These networks, foundational to modern deep learning, are trained by adjusting connection weights to minimize error, a process that relies on backpropagation to calculate gradients whenever the model is differentiable [1][3]. The new research introduces a unified vectorized feedback framework for learning on computational graphs, positioning synthetic gradients as a natural alternative [2]. The authors characterize the conditions under which synthetic gradients can achieve a lower gradient-estimation mean squared error than backpropagation [2]. They further construct examples showing that this sample efficiency advantage can be arbitrarily large [2]. Experiments on contextual bandits and reinforcement learning tasks demonstrate the potential of these theoretical findings [2]. Deep learning, which uses multilayered networks ranging from three to hundreds of layers, has driven advances in computer vision, speech recognition, and natural language processing since a resurgence in the 2010s fueled by GPU acceleration [3][5]. The field has since seen architectural innovations such as convolutional and transformer networks, the latter enabling large language models through attention mechanisms [4]. The proposal of synthetic gradients arrives amid an ongoing AI boom, where improving the efficiency of training algorithms is a central research concern [5].