Whale Speech: Approaching An Unknown Communication System By Latent Space Exploration And Causal Inference

Whale Speech: Approaching An Unknown Communication System By Latent Space Exploration And Causal Inference

Overview of the fiwGAN structure (Beguš, 2021b) and knowledge utilized in coaching. The determine illustrates three networks: the Generator with 5 upconvolutional layers, the Q-network with 5 convolutional layers, and the Discriminator with 5 convolutional layers. — stat.ML

This paper proposes a strategy for locating significant properties in knowledge by exploring the latent area of unsupervised deep generative fashions. We mix manipulation of particular person latent variables to excessive values exterior the coaching vary with strategies impressed by causal inference into an strategy we name causal disentanglement with excessive values (CDEV) and present that this strategy yields insights for mannequin interpretability. Utilizing this system, we will infer what properties of unknown knowledge the mannequin encodes as significant.

We apply the methodology to check what’s significant within the communication system of sperm whales, one of the intriguing and understudied animal communication programs. We practice a community that has been proven to be taught significant representations of speech and check whether or not we will leverage such unsupervised studying to decipher the properties of one other vocal communication system for which we’ve no floor reality.

The proposed approach means that sperm whales encode data utilizing the variety of clicks in a sequence, the regularity of their timing, and audio properties such because the spectral imply and the acoustic regularity of the sequences. A few of these findings are per present hypotheses, whereas others are proposed for the primary time.

We additionally argue that our fashions uncover guidelines that govern the construction of communication items within the sperm whale communication system and apply them whereas producing modern knowledge not proven throughout coaching.

This paper means that an interpretation of the outputs of deep neural networks with causal methodology is usually a viable technique for approaching knowledge about which little is thought and presents one other case of how deep studying can restrict the speculation area.

Lastly, the proposed strategy combining latent area manipulation and causal inference might be prolonged to different architectures and arbitrary datasets.

Gašper Beguš, Andrej Leban, Shane Gero

Feedback: 25 pages, 23 figures
Topics: Machine Studying (stat.ML); Machine Studying (cs.LG); Sound (cs.SD); Audio and Speech Processing (eess.AS)
Cite as: arXiv:2303.10931 [stat.ML] (or arXiv:2303.10931v1 [stat.ML] for this model)
Focus to be taught extra
Submission historical past
From: Andrej Leban
[v1] Mon, 20 Mar 2023 08:09:13 UTC (8,148 KB)
Astrobiology, SETI, CETI,

Explorers Membership Fellow, ex-NASA Area Station Payload supervisor/area biologist, Away Groups, Journalist, Lapsed climber, Synaesthete, Na’Vi-Jedi-Freman-Buddhist-mix, ASL, Devon Island and Everest Base Camp veteran, (he/him) 🖖🏻


Leave a Reply

Your email address will not be published. Required fields are marked *