📡 Signal & Communications¶

🧪 ICML2026 · 2 paper notes

Joint Model and Data Sparsification via the Marginal Likelihood: JMDS achieves simultaneous model and data sparsification through a unified objective of maximizing marginal likelihood. By avoiding the sub-optimality of multi-stage pipes, it maintains performance superior to independent sparsification across CIFAR, ImageNet, and WikiText at 5-10× joint compression ratios.
Meta-learning Structure-Preserving Dynamics: This paper systematically introduces modulation-based meta-learning (where a hyper-network maps latent codes \(\bm{z}^{(k)}\) to hierarchical modulation parameters) into Hamiltonian and GENERIC neural networks. It proposes two novel modulation schemes—latent multi-rank (MR) and latent SVD-like modulation—enabling a shared network to adapt to entire families of new parameter instances \(\bm{\mu}\) with few shots, while strictly maintaining energy conservation or dissipation structures.