MERL - Mitsubishi Electric Research Laboratories

TR2022-092

Disentangled surrogate task learning for improved domain generalization in unsupervised anomolous sound detection

    •  Venkatesh, S., Wichern, G., Subramanian, A.S., Le Roux, J., "Disentangled surrogate task learning for improved domain generalization in unsupervised anomolous sound detection," Tech. Rep. TR2022-092, DCASE2022 Challenge, July 2022.
      BibTeX TR2022-092 PDF
      • @techreport{Venkatesh2022jul,
      • author = {Venkatesh, Satvik and Wichern, Gordon and Subramanian, Aswin Shanmugam and Le Roux, Jonathan},
      • title = {Disentangled surrogate task learning for improved domain generalization in unsupervised anomolous sound detection},
      • institution = {DCASE2022 Challenge},
      • year = 2022,
      • month = jul,
      • url = {https://www.merl.com/publications/TR2022-092}
      • }
  • MERL Contacts:
  • Research Areas:

    Artificial Intelligence, Speech & Audio

Abstract:

We present our submission to the DCASE2022 Challenge Task 2, which focuses on domain generalization for anomalous sound detection. We investigated a novel multitask learning framework that disentangles domain shared features and domain-specific features. Disentanglement leads to better latent features and also increases flexibility in post-processing due to the availability of multiple embedding spaces. Our disentangled model obtains an overall harmonic mean of 74.57% on the development set, surpassing the MobileNetV2 baseline, which obtains 56.01%. Lastly, we explore the use of machine-specific loss functions and domain generalization methods, which improves our overall performance to 76.42%.