The NeurIPS 2025 paper “Auto-Compressing Networks” by Evangelos Dorovatas, Georgios Paraskevopoulos and Alexandros Potamianos investigates how network connectivity can encourage computational efficiency.
The proposed architecture uses long feedforward connections to promote useful representations in earlier layers. The paper examines the resulting compression behaviour, generalisation and robustness across several neural architectures.
This research connects the group’s machine learning activity with the practical challenge of reducing redundant computation.
- Authors: Evangelos Dorovatas, Georgios Paraskevopoulos, Alexandros Potamianos
- Venue: Advances in Neural Information Processing Systems 38, NeurIPS 2025
You may also like
-
ICLR 2026: measuring the hidden geometry of learned representations
-
TMLR 2026: prescriptive SVD-inspired attention via spectral energy retention
-
Towards Heritage World Models: from digital twins to predictive heritage systems
-
ICASSP 2026: representation-diverse self-supervision for bioacoustic learning
-
AISTATS 2026: process-tensor tomography of SGD and training memory