Learning dynamics of high-dimensional online ICA with moment-controlled non-Gaussian Data
Öz
Context—Many machine learning and signal processing methods assume Gaussian data because it simplifies analysis and often works reasonably well in practice. But real data are rarely Gaussian. They can have heavy tails, skewness, or unusually large kurtosis, and these differences matter. In fact, non-Gaussianity is exactly what makes problems like Independent Component Analysis (ICA) identifiable in the first place. What’s less clear is how these distributional differences affect learning dynamics, particularly in high-dimensional online settings. Recent work has shown that online ICA can be described using low-dimensional deterministic equations in certain scaling limits. However, most of these analyses assume a fixed source distribution and do not explore what happens when higher-order moments vary systematically. As a result, we still do not fully understand how changes in kurtosis or tail behavior influence stability, convergence speed, or sensitivity to the learning rate.
Objective—This work quantifies how controlled changes in higher-order moment structure shape the macroscopic dynamics of high-dimensional online ICA. We analyze how moment variations interact with initialization and learning rate to determine stability, convergence, and learning speed.
Method—We analyze a high-dimensional online ICA model through a McKean-type scaling limit, which yields a deterministic measure-valued evolution and an associated closed ordinary differential equation (ODE) for the alignment order parameter. We derive this macroscopic characterization for a broad class of nonlinear contrast functions and specialize the explicit stability and phase-transition analysis to the cubic nonlinearity. We then introduce a moment-controlled source model, constructed as a weighted combination of two non-Gaussian random variables, that preserves zero mean and unit variance while continuously tuning the fourth and sixth moments through a single parameter. In the cubic case, the limiting coefficients depend explicitly on these moments, allowing a detailed phase-plane and stability analysis. We validate the theoretical predictions by comparing the limiting ODE with simulations of the corresponding finite-dimensional online algorithm.
Results—The dynamics exhibit moment-induced phase transitions. Larger fourth and sixth moments shrink the basin of attraction of informative solutions, raise initialization thresholds, and reduce the admissible learning-rate range. In the cubic case, the signal-induced drift is governed by excess kurtosis, while the stochastic-damping contribution grows with higher-order moments, producing a trade-off between statistical signal strength and algorithmic stability. Simulations confirm slower convergence, increased sensitivity to initialization, and reduced robustness in large-moment regimes.
Conclusion—Non-Gaussianity is necessary for identifiability but can destabilize high-dimensional online learning. Our analysis reveals a fundamental trade-off between statistical richness and algorithmic stability in online ICA. The proposed framework directly links higher-order source statistics to learning dynamics and provides principled insight into learning-rate selection, initialization requirements, and extensions to alternative nonlinearities and multi-source settings.
Anahtar Kelimeler
- high dimensional asymptotics
- Independent Component Analysis
- non-Gaussian data
- online learning
- ordinary differential equations
Proje Numarası
Etik Beyan
Kaynakça
- S. Åberg, K. Podgórski, “A class of non-Gaussian second order random fields”, Extremes, 14(2), 187–222, 2011. https://doi.org/10.1007/s10687-010-0119-1.
- T. W. Lee, M. S. Lewicki, T. J. Sejnowski, “ICA mixture models for unsupervised classification of non-Gaussian classes and automatic context switching in blind signal separation”, IEEE Transactions on Pattern Analysis and Machine Intelligence, 22(10), 1078–1089, 2000. https://doi.org/10.1109/34.879789.
- A. Hyvärinen, E. Oja, “Independent component analysis: Algorithms and applications”, Neural Networks, 13(4–5), 411–430, 2000. https://doi.org/10.1016/S0893-6080(00)00026-5.
- T. H. Nguyen, U. Şimşekli, M. Gürbüzbalaban, G. Richard, “First Exit Time Analysis of Stochastic Gradient Descent Under Heavy-Tailed Gradient Noise”, Proceedings of the 33rd Conference on Neural Information Processing Systems (NeurIPS), Vancouver, Canada, 8–14 December 2019.
- C. J. Li, Z. Wang, H. Liu, “Online ICA: Understanding Global Dynamics of Nonconvex Optimization via Diffusion Processes”, Proceedings of the 30th Conference on Neural Information Processing Systems (NIPS), Barcelona, Spain, 5–11 December 2016, 4967–4975.
- Z. Jiao, M. Keller-Ressel, “Emergence of heavy tails in homogenized stochastic gradient descent”, Proceedings of the 38th Conference on Neural Information Processing Systems (NeurIPS 2024), Vancouver, Canada, 10–15 December 2024, 14066–14092. https://doi.org/10.52202/079017-0450.
- M. S. Advani, A. M. Saxe, H. Sompolinsky, “High-dimensional dynamics of generalization error in neural networks”, Neural Networks, 132, 428–446, 2020. https://doi.org/10.1016/j.neunet.2020.08.022.
- Z. Jin, B. B. Risk, D. S. Matteson, “Optimization and testing in linear non-Gaussian component analysis”, Statistical Analysis and Data Mining, 12(3), 141–156, 2019. https://doi.org/10.1002/sam.11403.
Ayrıntılar
Birincil Dil
İngilizce
Konular
Makine Öğrenmesi Algoritmaları
Bölüm
Araştırma Makalesi
Yazarlar
Zafer Doğan
*
0000-0002-5078-4590
Türkiye
Erken Görünüm Tarihi
14 Ağustos 2026
Yayımlanma Tarihi
-
Gönderilme Tarihi
7 Şubat 2026
Kabul Tarihi
31 Temmuz 2026
Yayımlandığı Sayı
Yıl 2026 Sayı: Advanced Online Publication