LooM: Self-Organizing Graph Autoencoders via Hebbian Learning LooM: Hebb Öǧrenmesi ile Kendini Örgütleyen Çizge Otokodlayicilar
34th Signal Processing and Communications Applications Conference, SIU 2026, İstanbul, Türkiye, 7 - 10 Temmuz 2026, (Tam Metin Bildiri)
- Yayın Türü: Bildiri / Tam Metin Bildiri
- Doi Numarası: 10.1109/siu71813.2026.11636636
- Basıldığı Şehir: İstanbul
- Basıldığı Ülke: Türkiye
- Anahtar Kelimeler: feature extraction, graph neural networks, Hebbian plasticity, self-organization
- Orta Doğu Teknik Üniversitesi Adresli: Evet
Özet
Most deep learning models rely on static, feedforward architectures defined prior to training, contrasting with the adaptive self-organization of biological systems. We introduce LooM, a self-organizing autoencoder where latent connectivity emerges through Hebbian plasticity. While traditional autoencoders treat latent variables as isolated vectors, LooM extracts feature maps from convolutional layers and organizes them into a dynamically evolving graph. By utilizing feature co-activation statistics as a teacher, the model learns adjacencies that reflect relationships between extracted features. Message passing over this learned graph generates a topologically-informed latent space. The resulting unified model integrates biologically inspired plasticity with neural networks, offering a novel approach to unsupervised feature orchestration and explainable latent representations.