Israel finds AI limitations in decoding animal signals — Jerusalem Post

Researchers at Tel Aviv University in Israel found that artificial intelligence models can make mistakes when attempting to decode animal communication because they analyze acoustic differences that are not necessarily meaningful to the signal’s recipient. The results of the study were published in the journal Current Biology, the Jerusalem Post reports.

The scientists tested this approach on vocalizations of young children who had not yet mastered full speech. The recordings covered three situations: distress, addressing a specific person — the mother or father — and requesting food. The authors used a classical acoustic analysis method and two modern deep neural networks: one trained on animal vocalizations and the other on adult human speech.

Models did not determine the meaning of sounds

The neural networks performed better than the classical method at grouping sounds by their characteristics, but failed to classify children’s vocalizations by meaning. In some cases, the systems grouped signals with different messages together, while in others they separated vocalizations that conveyed the same meaning.

More current news is available on the UA.News Telegram channel Telegram.

The models also failed to recognize a sequence of sounds that expressed increasing urgency. People who listened to these recordings perceived the growing urgency through subtle acoustic changes.

Response and context are important

According to the researchers, two sounds that differ significantly on a spectrogram may carry the same message for the recipient. At the same time, acoustically similar signals may have different meanings depending on the context. Therefore, classification based solely on sound similarity can create a distorted picture of a communication system.

The study’s lead author, Yossi Yovel, noted that to understand what an animal is communicating, it is necessary to know how the recipient animal perceives and responds to the signal. In his view, reliable decoding of animal communication requires combining AI with behavioral observations, signal playback experiments and, in some cases, measurements of brain activity.