Deep learning for ethoacoustical mapping: Application to a single Cachalot long term recording on joint observatories in Vancouver Island

Abstract

During February and March, 2018, a lone sperm whale known as Yukusam was recorded first by Orcalab in Johnstone Strait and subsequently on multiple hydrophones within the Salish Sea [1]. We learn and denoise these multichannel clicks trains with AutoEncoders Convolutional Neural Net (CNN). Then, we build a map of the echolocations to elucidate variations in the acoustic behavior of this unique animal over time, in different environments and distinct levels of boat noise. If CNN approximates an optimal kernel decomposition, it requires large amounts of data. Via spline functionals we offer analytics kernels with learnable coefficients do reduce it. We [1-3] identify the analytical mother wavelet to represent the input signal to directly learn the wavelet support from scratch by gradient descend on the parameters of cubic splines [2]. Supplemental material http://sabiod.org/yukusam [1] Balestriero, Roger, Glotin, Baraniuk, Semi-Supervised Learning via New Deep Network Inversion, arXiv preprint arXiv:1711.04313, 2017 [2] Balestriero, Cosentino, Glotin, Baraniuk, WaveletNet : Spline Filters for End-to-End Deep Learning, Int. Conf. on MachineLearning, ICML, Stockholm, http://sabiod.org/bib, 2018 [3] Spong P., Symonds H., et al., Joint Observatories Following a Single male Cachalot during 12 weeks—The Yukusam story, ASA 2018.

Toward understanding the communication in sperm whales

“…Given the large magnitude of data, a key step is to build appropriate data storage and processing infrastructure, including automated ML pipelines (maintainable and reusable across multiple data collecting devices) that will replace the annotation currently done largely by hand by marine biologists. ML-based methods are already being used for detection and classification among marine mammals ( Gillespie et al., 2009 ; Shiu et al., 2020 ) and for sperm whale click detection and classification ( Bermant et al., 2019 ; Ferrari et al., 2020 ; Glotin et al., 2018 ; Jiang et al., 2018 ); such methods are potentially scalable to large datasets containing years of recording that would otherwise be beyond reach with previous manual approaches.…”