Efficient Learning in Predictive Coding Networks Using Global Error Signals

The error backpropagation is a representative learning algorithm widely adopted across deep learning networks. Despite its success in achieving high learning performance and generalization, error backpropagation has been criticized for its biological implausibility. One of its primary limitations lies in the weight transport problem, where, unlike the human brain, error backpropagation requires the same weights used in forward propagation to be accessible in the backward propagation. In addition, the sequential calculation of error backpropagation restricts parallelization that results in reduced efficiency. To solve these problems, predictive coding networks have emerged as an alternative. Inspired by predictive coding in computational neuroscience, it offers a biologically plausible learning framework. However, as network depth increases, predictive coding can suffer from significant computational costs, which decreases efficiency. In this paper, we propose an approach that enhances the efficiency of predictive coding by using global error signals. We also evaluate the effectiveness of our method on benchmark datasets, demonstrating improved learning performance and computational efficiency.