TinyCast is an attention-free, 146,505-parameter zero-shot forecaster that uses a zero-parameter spectral detector and dilated convolutional encoder. It defines the size-accuracy frontier for small models by computing periodic structures rather than learning them.
HOW THIS AFFECTS YOU
●
builderYou can deploy extremely lightweight, sub-1M parameter models for probabilistic forecasting on edge devices.
●
researcherThis provides a highly efficient alternative to attention-based architectures for periodic time-series tasks.