Meta-AF: Meta-Learning for Adaptive Filters

Adaptive filtering algorithms are pervasive throughout signal processing and have had a material impact on wide variety of domains including audio processing, telecommunications, biomedical sensing, astrophysics cosmology, seismology, many more. filters typically operate via specialized online, iterative optimization methods such as least-mean squares or recursive least aim to process signals in unknown nonstationary environments. Such algorithms, however, can be slow laborious develop, require domain expertise create, necessitate mathematical insight for improvement. In this work, we seek improve upon hand-derived adaptive filter present comprehensive framework learning update rules directly from data. To do so, frame the development meta-learning problem context deep use form self-supervision learn online filters. demonstrate our approach, focus applications systematically develop meta-learned five canonical problems system identification, acoustic echo cancellation, blind equalization, multi-channel dereverberation, beamforming.We compare approach against common baselines and/or recent state-of-the-art methods. We show high-performing that real-time and, most cases, significantly outperform each method – all using single general-purpose configuration approach.