Meta-learning for adaptive filtering

Adaptive filtering algorithms form a cornerstone of intelligent signal processing infrastructure and play a vital role in a wide array of applications, including telecommunications, biomedical sensing, and seismology. Adaptive filters typically operate via specialized, online, iterative optimization methods such as least-mean squares or recursive least squares and aim to process signals in unknown or nonstationary environments. Such algorithms, however, can be slow and laborious to develop, require domain expertise to create, and necessitate mathematical insight for improvement. In this thesis, we present a different approach to the design of adaptive filters. Our core contribution is a comprehensive framework that leverages the power of meta-learning and deep learning techniques to learn adaptive signal processing algorithms directly from data. By formulating adaptive filter development as a meta-learning problem, we train iterative update rules for adaptive filters using various forms of supervision and training. This approach enables us to overcome the limitations of traditional design methodologies and opens up new possibilities for developing highly efficient and effective adaptive filters. To validate our framework, we focus on audio applications and systematically develop a family of meta-learned adaptive filters for key tasks such as system identification, inverse modeling, prediction, and informed interference cancellation. Through extensive evaluations of diverse audio applications, including acoustic echo cancellation, blind equalization, multi-channel dereverberation, beamforming, telephony, and keyword spotting, we demonstrate the strong performance of our approach compared to existing methods. Our findings indicate that our meta-learned adaptive filters not only function in real-time but also consistently excel across different tasks. We achieve remarkable performance gains by employing a single general-purpose configuration, highlighting the versatility and effectiveness of our framework. Moreover, our work pushes the boundaries of meta-learning and adaptive filter literature, leading to a new conceptual framework that has the potential to change how we approach, solve, and construct adaptive filter pipelines.

Name thesis:degree_name

Ph.D.

Level thesis:degree_level

Dissertation

Discipline thesis:degree_discipline

Computer Science

Grantor

University of Illinois at Urbana-Champaign

Year dc:date

2023

Author dc:creator

• Casebeer, Jonah

Contributors dc:contributor

• Smaragdis, Paris
• Singer, Andrew
• Vasisht, Deepak
• Soltanaghai, Elahe
• Bryan, Nicholas J

• Adaptive Filtering
• Meta-learning
• Online Optimization
• Learning To Learn
• Deep Learning
• Machine Learning
• Signal Processing

Statement dc:rights

• Copyright 2023 Jonah Casebeer

Language dc:language

en, eng

Handle dc:identifier

https://hdl.handle.net/2142/121970

OAI record oai:identifier

https://www.ideals.illinois.edu/oai-pmh?verb=GetRecord&metadataPrefix=etdms&identifier=oai%3Awww.ideals.illinois.edu%3A2142%2F121970