A recent arXiv paper explores weight perturbation (WP) as an alternative to backpropagation for training neural networks. WP estimates learning directions by applying random perturbations to parameters and observing the resulting scalar change in loss, without computing derivatives via backpropagation. The authors focus on the MNIST dataset, a standard benchmark for image classification.
The abstract outlines an investigation into how batch size, training duration, and estimator stabilization affect WP's performance. These factors are likely critical for making WP competitive with backpropagation, which remains the dominant training method. The paper's goal, as stated in its title, is to approach backpropagation-level accuracy using WP.
Because only the abstract is available, specific numerical results or conclusions are not yet reported. The study appears to be a systematic exploration of WP's practical viability, rather than a proposal of a new algorithm. The findings could inform future work on biologically plausible learning rules or hardware-friendly training methods.