Analysis
We evaluated Perch 2.0 utilizing a few-shot linear probe on marine duties, akin to distinguishing completely different baleen whale species or completely different killer whale subpopulations. Its efficiency was in contrast in opposition to pre-trained fashions which can be supported in our Perch Hoplite repository for agile modeling and switch studying. They embody Perch 2.0, Perch 1.0, SurfPerch, and the multispecies whale mannequin.
For underwater knowledge analysis, we used three datasets: NOAA PIPAN, ReefSet, and DCLDE.
- NOAA PIPAN: An annotated subset of the NOAA NCEI Passive Acoustic Information Archive from the NOAA Pacific Islands Fisheries Science Heart recordings. It consists of labels utilized in our prior whale fashions in addition to new annotations for baleen species akin to frequent minke whale, humpback whale, sei whale, blue whale, fin whale, and Bryde’s whale.
- Reef Set: Developed for SurfPerch mannequin coaching, this dataset leverages knowledge annotations from the Google Arts and Tradition mission: Calling in Our Corals. It consists of a mixture of organic reef noises (croaks, crackles, growls), particular species/genera courses (e.g., damselfish, dolphins, and groupers), and anthropomorphic noise and wave courses.
- DCLDE: This dataset is evaluated utilizing three completely different label units:
- Species: For distinguishing between killer whales, humpbacks, abiotic sounds, and unknown underwater sounds (with some uncertainty in killer whale and humpbacks labels).
- Species Recognized Bio: For sure labels of killer whales and humpbacks.
- Ecotype: For distinguishing between killer whale subpopulations (ecotypes), together with Transient/Biggs, Northern Residents, Southern Residents, Southeastern Alaska killer whales, and offshore killer whales.
On this protocol, for a given goal dataset with labeled knowledge, we compute embeddings from every of the candidate fashions. We then choose a hard and fast variety of examples per class (4, 8, 16, or 32), and prepare a easy multi-class logistic regression mannequin on high of the embeddings. We use the ensuing classifier to compute the realm underneath the receiver-operating attribute curve (AUC_ROC), the place values nearer to 1 point out a stronger capability to tell apart between courses. This course of simulates utilizing a given pre-trained embedding mannequin to create a customized classifier from a small variety of labelled examples.
Our outcomes present that extra examples per class enhance efficiency throughout all of the fashions, besides on ReefSet knowledge, the place efficiency is excessive even with solely 4 examples per class for all fashions, besides the multispecies whale mannequin. Notably, Perch 2.0 is constantly both the highest or second-best performing mannequin for every dataset and pattern dimension.
