Kuromi: Learning without Augmentation via Energy-based Semi-supervised Kuramoto Neurons

Hongyang He, Xinyuan Song, Hongyang Xie, Yan Zhong, Victor Sanchez

Keywords: Semi-supervised Learning

TL;DR: A semi-supervised learning framework without data augmentation

Abstract: Semi-supervised learning (SSL) often relies on extensive data augmentation or complex teacher-student structures to generate reliable supervision from unlabeled data. In this work, we propose Kuromi, a novel SSL framework that discards data-space augmentation entirely and instead leverages energy-based dynamics for representation learning and pseudo-label generation. Central to Kuromi is the Artificial Kuramoto Oscillatory Neuron (AKOrN), a biologically inspired dynamic neuron model that encodes inputs as synchronized oscillatory states on a hypersphere. Through unsupervised pre-training, fine-tuning on limited labels, and energy-guided self-training, Kuromi produces low-energy, structure-aware feature representations without requiring external regularization. To better exploit labeled data and pseudo-labels, we also propose Energize, an iterative, augmentation-free latent-state smoothing method inspired by Weisfeiler-Lehman aggregation, which operates entirely in the prediction space without hyperparameters. Extensive experiments on CIFAR-10, SVHN, STL-10, and ImageNet show that Kuromi achieves state-of-the-art performance among non-transformer SSL methods and remains competitive with recent transformer-based approaches. Notably, Kuromi not only achieves state-of-the-art results on CIFAR-10 and STL-10, but also leads in efficiency with just 12M parameters, 0.18 GFLOPs, and the highest throughput across all baselines.

Primary Area: unsupervised, self-supervised, semi-supervised, and supervised representation learning

Submission Number: 1051