Tagungsband - DAGA 2024
Beitrag
Efficient Scene Recognition with Sparse Butterfly Student Training (no manuscript)
Zusammenfassung:
Acoustic Scene Classification (ASC) sets high complexity constraints on model architectures for real-world deployment. In this talk, we apply recent advances in linear operator factorization and sparsification to the ASC task. Inspired from signal processing, Butterfly matrices allow sub-quadratic parameter growth and make smaller Transformer models possible. We propose additionally probabilistic subsampling attention based on the Straight-Through Gumbel estimator. The probabilistic sparsification map makes a student-teacher training adaptive to Multiply-Accumulate (MAC) and parameter count constraints. We demonstrate strong performance on the “TAU Urban Acoustic Scenes 2020 Mobile" dataset under different subsampling rates in frequency-time domain and compare to recently proposed CP-ResNet models.