Szczegóły publikacji
Opis bibliograficzny
The impact of optimiser choice on the training dynamics of complex-valued neural networks: a comparative study with real-valued counterparts / Michał RUTKOWSKI, Piotr A. KOWALSKI // International Journal of Approximate Reasoning ; ISSN 0888-613X . — 2026 — vol. 197 art. no. 109722, s. 1–18. — Bibliogr. s. 17–18, Abstr. — Publikacja dostępna online od: 2026-05-22. — P. A. Kowalski - dod. afiliacja: Systems Research Institute, Polish Academy of Sciences, Warsaw
Autorzy (2)
Słowa kluczowe
Dane bibliometryczne
| ID BaDAP | 169723 |
|---|---|
| Data dodania do BaDAP | 2026-09-02 |
| Tekst źródłowy | URL |
| DOI | 10.1016/j.ijar.2026.109722 |
| Rok publikacji | 2026 |
| Typ publikacji | artykuł w czasopiśmie |
| Otwarty dostęp |  |
| Czasopismo/seria | International Journal of Approximate Reasoning |
Abstract
Complex-valued neural networks (CVNNs) are often credited with advantages on signals that carry phase and amplitude, yet empirical comparisons with real-valued neural networks (RVNNs) are frequently confounded by architectural and capacity mismatches – and by the underexplored role of the optimiser. We present a controlled study that isolates the effect of optimiser choice on the training dynamics and generalisation of CVNNs across two benchmark settings: MNIST image classification and the SynSinSig-10 synthetic time-series dataset. The former serves as a widely used controlled benchmark for optimisation analysis, whereas the latter provides a signal-oriented setting in which discriminative structure is directly related to frequency and phase. In each case, we compare complex-valued multilayer perceptrons using ZERO and Fourier-domain input representations (FFT2 for MNIST, FFT for SynSinSig-10) against real-valued surrogates—topology- and parameter-matched—constructed via the REAL, TWICE, and PARAM mappings. Four widely used optimisers—SGD, Adagrad, RMSprop, and Adam—are evaluated across multiple seeds, and we analyse full loss trajectories, minima, and generalisation gaps. Across both datasets, the results reveal a pronounced interaction between optimiser choice, input representation, and task characteristics. Fourier-based complex representations are the most optimiser-sensitive: they can be highly effective under SGD, but are markedly less robust under adaptive methods. By contrast, ZERO and the real-valued baselines are more stable, although their relative ranking remains problem-dependent. On MNIST, the capacity-aware mappings (TWICE, PARAM) reliably outperform a naïve one-to-one replacement (REAL), whereas on SynSinSig-10 the ranking among real-valued mappings becomes more task- and optimiser-dependent. Overall, CVNNs are not uniformly superior to RVNNs; their advantages materialise when the optimiser and representation are aligned. Beyond headline accuracy, our analysis clarifies when complex-aware treatment is warranted and provides a principled recipe for constructing fair RV baselines, enabling more meaningful CVNN-RVNN comparisons.
