At this point, I am well aware of the ubiquity of Artificial Intelligence and Machine Learning in our lifes today, and that living without them would set us back decades. One of the great things it allows us to access is a huge diversity of languages we do not necessarily speak, thanks to a variety of online dictionaries and translators that can convert our spoken or written language into another one, even if these translations sometimes turn out to be rather… interesting.
However, one thing I had never heard about before I recently stumbled across an article about it was the use of machine learning for the translation of animal language. Honestly, to me, this sounded more like the plot of a Sci-Fi movie, but there are indeed several research projects running that are attempting to do that exact thing. And while the achievement of their final goal still seems to lie in the more distant future, they have had some success that I would like to tell you about. As stated before, there are several groups actively working on the idea of translating animal language, but the most ambitious one is the nonprofit Earth Species Project (ESP). This California-based group aims to decode communication of all non-human species and subsequently make these discoveries available to the public to help us understand them, in the hopes of deepening the bond between humans and nature and thus helping protect them. According to them, a lot of our maltreatment of animals results from issues in communication, and an attempt at translation might deepen our respect for them and relieve us of our human preconceptions. The biggest obstacle in this project is the mere fact that, as opposed to other researchers, they wish to include every species, especially because their ways of non-verbal communication differ immensely from each other.
This huge project is supposed to work thanks to the development of an algorithm that would geometrically represent the words in a physical space, in which the direction and distance between different words stand for their semantic relationship to each other. The example given here is the following: king has to man the same distance and direction that woman has to queen. The mapping of these points is made merely by observing how often they appear near one another, without necessarily knowing their meaning. The shapes emerging from this process are comparable for different languages and would thus allow a translation through their alignment.
So, how likely is it that they will succeed? To be fair, my understanding of this topic is very limited, but so far, they have reached two milestones that pushed them a few steps toward their target.
Imagine being in a really busy bar, talking to a friend. The music is quite loud, and everyone else in the room is engaged in their own conversation, raising their voices to fight against the background noise. To be able to keep up your conversation you will need to pick out that specific sound that is of interest to you, in this case, your friend’s voice, and ignore all the other sounds.
Press play to be transported to a noisy cocktail party. At first, you can only hear a melange of sounds. Then, you run into an old friend, who starts talking to you. As you focus on what your friend is saying, you are eventually able to filter out the other sounds of the party, effectively turning them into background noise.
This phenomenon is called the ‘cocktail party problem’, and ESP managed to solve it in their research. They created a code that can make out which animal out of a big group of noisy animals is vocalizing. Another thing they have figured out is how to ‘talk’ to animals directly. The group came up with an algorithm that enables the AI to generate mimicked calls, the meaning of which are still a mystery to us.
In a project that is currently still in progress, ESP is working on the development of an algorithm that can determine the number of call types one species uses. In order to do so, they are applying self-monitored ML to learn patterns. Another project is concerned with the operative meanings of vocalizations and their automatic understanding. For this, they are studying the behavior of wild sea mammals underwater with the use of small biologging devices in one of the largest tagging programs in the world in which they record the animals’ location, motion, and visual field. The aim here is to automatically measure their activities and add audio data to understand whether the sounds tied to their behavior have a specific functional meaning.
I am always a little critical when it comes to projects like these, and of course, I am not the only one. While the main reason for the group to carry out this lengthy research is that of the creation of kinship with nature and damage control, it is always extremely important to consider the way that these advancements will be used. Critics of projects like ESP generally highlight the risk of the abuse of our newfound powers to dominate the natural world of animals and plants, exploiting them, and carrying on the military history of using animals.
What do you think of this project?
ESP is certainly not the only group actively working on decoding animal language, but their project is the most extensive one. Other groups that might be of interest to you, if you wish to further read up on this topic would be the Wild Dolphin Project, the Interspecies internet, DeepSqueak, and Project CETI (Cetacean Translation Initiative).
Charlotte Hu, “Artificial intelligence is helping scientists decode animal languages: A Google translate for rodents and whales doesn’t exist yet, but researchers are working on it,” Popular Science, 1st September 2022.