Scientists are using artificial intelligence to translate animal languages.

A cartoon dog named Dug wears a kind of magical collar in the Pixar film Up that can convert his barks and whines into fluid human speech. Very well-trained canines can be taught to press buttons that play human voice when given simple commands like "outside," "walk," and "play" elsewhere in the real world. The idea of being able to communicate with the animals they coexist with has always interested humans, and recently, machine learning has emerged as a promising path toward animal translation thanks to its always improving capabilities for understanding human speech.

This week, a big effort by five research teams was described in a New York Times article. These teams looked into utilizing machine-learning algorithms to evaluate the sounds of rodents, lemurs, whales, chickens, pigs, bats, cats, and other animals.

Artificial intelligence systems typically acquire new skills by training with labeled data (which can be supplied by the internet, or resources like e-books). For computer programs that simulate human language, this typically entails giving them a sentence, leaving out some words, and asking the program to complete the sentence. Additionally, there are more original approaches being used now to link speech to brain activity.

But examining animal language differs from examining only human language. Software must be taught by computer scientists what to look for and how to organize the data. The majority of the time, this technique depends on matching vocal recordings with the visual social activities of animals in addition to amassing a large number of vocal recordings. For instance, a team researching Egyptian fruit bats recorded the bats using video cameras to give the sound context. Additionally, the team studying whales intends to use audio, video, and tags that can capture animal movements in order to understand the syntax and semantics of what whales are saying and why.

For the better part of the past ten years, there has been an idealistic initiative to create a Google Translate for animals. The ability of machine learning to recognize animals by call and even detect their presence has advanced significantly. (The Merlin app from Cornell is amazingly accurate in identifying different bird species from their cries.) Even though this kind of software has demonstrated some success in identifying the fundamental vocabulary of some animals from the characteristics of their vocalizations (i.e. frequency or loudness) as well as attributing calls to specific individuals, it is still a long way from fully grasping all the nuanced nuances that animal language may encompass.

Many opponents of this strategy point to the inadequacies of current AI language models in terms of their capacity to comprehend the connections between words and the things they may refer to in the real world, as well as the limitations of scientists' understanding of animal societies in general. The basis of artificial intelligence language models for people is a computer model of the link between words and possible situations (where they might go in a sentence, and what they might refer to). Although researchers are aware of what goes in and what comes out, these models sometimes operate like a "black box," leaving them unsure of how the algorithm arrived at the result.

The idea that animal communications may not function at all like human communications and that the propensity to anthropomorphize them may be skewing the results is another consideration for researchers. Due to physiologic and behavioural variations, animal language may contain special components.

According to a Wall Street Journal report earlier this year, there are proposals for using self-supervised learning algorithms to analyse audio data in order to get around the inability to know the data parameters in advance. In these algorithms, the computer informs the researchers of any patterns it discovers in the data—patterns that could reveal connections that the human eye would have missed. Human goals for this kind of research ultimately determine how deep humans travel down the rabbit hole of attempting to understand animal interactions, and for those purposes it may be sufficient to grasp the fundamentals.