Learning What Remains: A Universal Framework for Multi-Domain Neural Decomposition

We introduce Learning What Remains (LWR), a framework where foundational architectures generate rep-resentations while specialized branches learn from what remains. Any architecture can act as a base or branch,employing sequential or parallel processing. We showcase LWR’s adaptability across various domains: achieving0.761 OWA on M4 time series (surpassing N-BEATS’ 0.795), and 87.80% ± 2.64 accuracy on ESC-50 envi-ronmental sound classification (approaching ERANNs’ 89.2% from scratch). The framework facilitates swiftexperimentation with architectural combinations without the need for domain-specific redesign.