in a pixar movie Up, a cartoon dog called Doug is equipped with a kind of magical collar that can convert his barking and whining into fluent human speech. Elsewhere in the real world, we ask very well-trained dogs to press buttons that generate human speech with simple commands like “get out,” “walk,” and “play.” I can teach. Humans have always been fascinated by the possibility of communicating with the animals they share their world with. Recently, machine learning with increasingly sophisticated capabilities for parsing human speech has been presented as a promising route to animal translation.

Article new york times This week, we documented a major effort by five research groups looking to use machine learning algorithms to analyze the sounds of rodents, lemurs, whales, chickens, pigs, bats, cats, and more.

Artificial intelligence systems typically learn through training using labeled data (provided by resources such as the internet or e-books). For human language models, this typically involves giving the computer a sentence, blocking certain words, and asking the program to fill in the blanks. There are also more creative strategies that try to match speech with brain activity.

But analyzing animal language is different from analyzing human language. Computer scientists need to tell software programs what to look for and how to organize data. This process almost always relies not only on accumulating a sufficient number of voice recordings, but also on matching these voice recordings with the visual social behavior of the animal. used video cameras to record the bats themselves to provide context for the call. Groups studying whales also use video, audio, and tags that can record animal movements to decipher the syntax, semantics, and ultimately what the whales are communicating and why. are planning Of course, some groups have also suggested testing animal dictionaries by playing recordings to animals and examining their responses.

Creating Google Translate for animals has been an ambitious project we’ve been working on for more than half the last decade. Machine learning has also made great strides in detecting the presence of animals and, in some cases, accurately identifying animals by their vocalizations. (Cornell’s Merlin app can accurately match bird species to their calls.) This kind of software can match a particular animal’s basic vocabulary to vocal characteristics (i.e., frequency or volume). have some success in identifying as well. While we can attribute vocalizations to individuals, we are still far from understanding all the complex nuances that animal language can encapsulate.

[Related: With new tags, researchers can track sharks into the inky depths of the ocean’s Twilight Zone]

Many skeptics of this approach argue that current AI language models have the shortcomings of being able to truly understand the relationship between words and the objects they may refer to in the real world, as well as the ability of scientists to understand animal societies as a whole. It points out both flaws in your understanding. Artificial intelligence language models for humans rely on computers mapping the relationship between words and the contexts in which they may appear (where they go in a sentence, what they refer to). However, these models have their own flaws and can be black boxes. Researchers know what goes in and what comes out, but they don’t fully understand how the algorithm reaches its conclusions.

Another factor the researchers are considering is the fact that animal communication may not function at all like human communication, and the tendency to anthropomorphize them may have distorted the results. I have. Due to physiological and behavioral differences, there may be unique elements in animal language.

For the purpose of not knowing the data parameters in advance, there have been proposals to analyze audio data using self-supervised learning algorithms. wall street journalThis allows the computer to tell researchers what patterns they see in the data, patterns that may reveal connections that the human eye misses. Ultimately, how far down the rabbit hole humans go in trying to understand animal communication depends on their goals for this kind of research, and for that purpose, understanding the basics may be enough. For example, a translator that can reliably interpret whether the animals we come into frequent contact with are happy, sad, or in danger would be useful and practical to create.