FEATHer, or the Fourier-Efficient Adaptive Temporal Hierarchy Forecaster, is a novel deep learning architecture specifically engineered for time-series forecasting in environments with severe computational and memory constraints, such as edge devices like PLCs and microcontrollers. Its core mechanism involves a unique combination of ultra-lightweight multiscale decomposition into frequency pathways, a shared Dense Temporal Kernel that avoids recurrence or attention, and frequency-aware branch gating for adaptive representation fusion. FEATHer also incorporates a Sparse Period Kernel to effectively capture seasonality through period-wise downsampling. This design addresses the critical problem of deploying accurate long-term forecasting models in industrial domains like manufacturing and smart factories, where conventional deep architectures are often impractical due to their large parameter counts. By maintaining a compact architecture, sometimes with as few as 400 parameters, FEATHer enables reliable forecasting directly on edge devices, outperforming baselines across various benchmarks.
FEATHer is a new, very small AI model designed to predict future trends in data, like factory sensor readings, directly on tiny computers. It's built to be super efficient, using only a few hundred parameters, so it can run fast and accurately on devices with very limited power and memory, which traditional large AI models cannot do.
Fourier-Efficient Adaptive Temporal Hierarchy Forecaster
Was this definition helpful?