Emeka Ndupuechi, Christian Beecks · Machine Learning with Applications 2026 · 2026
DOI: 10.1016/j.mlwa.2026.101059
Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).
Convolutional neural networks (CNNs) primarily rely on spatial-domain feature processing, where convolutional feature maps are computed using learned kernel weights over local spatial neighbourhoods. To complement this local spatial processing, we propose LiteFA-Net, a lightweight frequency-adaptive CNN that integrates frequency-domain conditioning into spatial feature learning. Specifically, LiteFA-Net integrates a spatial backbone (Lite-Net) with a novel frequency-adaptive enhancement framework that leverages globally-averaged spectral magnitudes from all spatial locations in the spatial feature map of each channel as conditioning signals to compute learned frequency-conditioned gating functions that refine early spatial feature maps, and modulate outputconvolutional feature maps and feature fusion. Through this frequency-conditioned gating mechanism, forward activations and backpropagated gradients are adaptively modulated, thereby conditioning spatial feature learning on channel-wise spectral information, which encourages more variability in gradient magnitudes, enhances feature learning adaptability, and improves generalization. Experimental results demonstrate that LiteFA-Net outperforms the best-performing CNN and transformer baselines across evaluated model-capacity regimes. Specifically, on CIFAR-10/CIFAR-100, LiteFA-Net achieves Top-1 testaccuracy gains of +3.50%/+4.14% (with 1.07 × more parameters and 1.25 × higher MACs), +1.07%/+2.94% (with 1.8 × fewer parameters and 1.3 × higher MACs), and +0.40%/+1.75% (with 1.09 × more parameters and 2.0 × fewer MACs) in the Lower, Intermediate, and Upper Model Regimes, respectively. On ImageNet-100, LiteFA-Net-S improves mean Top-1 test accuracy over CCT-7/3 × 1 by +2.23% (with 1.05 × more parameters and 3.1 × fewer MACs), and over FFC-ResNet-50 by +10.20% (with 6.8 × fewer parameters and 6.2 × higher MACs). Code and results: https://github.com/Ndupuechi/LiteFA-Net .
No comments yet — start the discussion below.