Frequency Selective Neural Networks Offer Interpretable Time Series Learning

A new deep learning architecture, the Frequency Selective Neural Network (FSNN), has been proposed as a foundational model for time series analysis. Unlike conventional deep learning models such as CNNs, RNNs, and Transformers, which often treat continuous waves as abstract spatial or token sequences, FSNN directly integrates advanced signal processing mathematics into its design. This approach aims to overcome "spectral entanglement," where predictive accuracy is decoupled from physical reality. FSNN utilizes a fully differentiable Wiener-like filter bank, optimized through complex-domain backpropagation, to autonomously identify and isolate the specific physical oscillatory modes relevant to a given task. This inherent design allows FSNN to achieve physical interpretability, meaning its outputs directly correspond to meaningful frequency bands, such as isolating the cardiac QRS complex in ECG data. Extensive evaluations demonstrate that FSNN achieves state-of-the-art predictive performance across multiple benchmarks, including multivariate UEA datasets and the PTB-XL clinical ECG benchmark. Crucially, it provides this superior performance while offering a highly scalable and interpretable paradigm for robust pattern recognition in complex temporal domains, addressing a long-standing challenge in deep learning for time series.