HAMON: Passive Optical Sequence Mixing for Long-Horizon Forecasting
Researchers have developed HAMON, a novel passive optical system for long-horizon time-series forecasting. This system encodes historical data onto an optical aperture and uses trainable phase masks with free-space diffraction to shape forecasts directly in the output field, eliminating the need for digital sequence-mixing layers during inference. HAMON has demonstrated superior performance on benchmarks like ETTm2 and ETTh2, outperforming strong digital baselines by up to 14% in MSE, while showing competitive results on weather data and mixed performance on traffic and electricity datasets. AI
IMPACT This passive optical approach could offer a new substrate for sequence mixing, potentially leading to more efficient forecasting models.