Year: 2026
Type: review
Abstract: This paper is a Japanese tutorial article that summarizes our replication of the Forward-Forward (FF) algorithm on MNIST in the supervised label-in-input setting. We verify the correspondence between the implementation and the mathematical formulation and organize our hyperparameter exploration in a hypothesisintervention-result-interpretation-next-step format, providing tables and figures aggregated across runs from the experiment logs. In our replication, the best test error under label-search inference was 2.12% with jitter augmentation (± 2 pixels) after 500 epochs.
Authors: Shirahama, Naruki, Nakaya, Naofumi, 渡邉, 志
Institutions: Shimonoseki City University, Juntendo University
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7163969920