REALM: Retrospective Encoder Alignment for LFP Modeling
REALM is a framework for high-performance, real-time decoding of brain activity from Local Field Potentials (LFPs) using a retrospective distillation pipeline. It trains a compact causal model by distilling knowledge from a larger, non-causal teacher model through representation alignment and task supervision. This approach achieves state-of-the-art performance with significantly reduced parameters and training time, enabling deployment on low-power edge devices.