동물의 의사소통과 언어 형태를 해독하는 AI

### AI로 동물의 의사소통 해독하기: 코끼리 '이름'부터 종간 언어 모델까지
* 🐘 조이스 풀과 미키 파르도 같은 연구자들은 머신러닝을 이용해 코끼리 울음소리를 분석하여 **아프리카 사바나 코끼리들이 서로 이름을 부른다는 증거**를 발견했습니다.
* 🔊 여러 동물이 시끄러운 환경에서 발성하는 동물 소리 녹음에서 **'칵테일 파티 문제'는 개별 소리를 불협화음에서 분리할 수 있는 AI 알고리즘**을 통해 해결되고 있습니다.
* 🦇 이집트 과일박쥐 발성에 대한 요시 요벨의 연구는 **지도 학습을 사용하여 발화자, 발화 맥락, 발화 수신자를 식별**했지만 인간 라벨이 붙은 학습 데이터의 한계를 강조했습니다.
* 🤖 인간 언어용 ChatGPT와 같은 자가 지도 학습 모델은 제한된 인간의 주석에 의존하지 않기 때문에 **동물의 의사소통을 해독하는 데 가장 큰 잠재력**을 가진 것으로 간주됩니다.
* 🌐 지구 종 프로젝트는 언어에는 서로 다른 언어 간에 정렬할 수 있는 '모양'이 있다는 아이디어를 바탕으로 **종 간 의사소통을 위한 대규모 언어 모델**을 구축하는 것을 목표로 합니다.
* 🐬 연구자들은 인간과 동물의 의사소통 모델을 정렬하면 **공통점을 발견**하고 비인간 지능과 의식에 대한 이해가 깊어질 것으로 기대하고 있습니다.
* 📊 이 과정의 첫 번째 단계는 동물 소리에 대한 광범위한 데이터를 수집하는 것으로, 아자 라스킨과 같은 연구자들은 **수천 개의 개별 통화로 데이터베이스**를 구축하기 위해 노력하고 있습니다.
* 🌍 AI가 촉진하는 진정한 종 간 소통은 아직 먼 목표이지만, 연구자들은 이 여정이 **보존 노력과 다른 종과의 관계에 영향을 미칠 수 있는 중요한 발견**으로 이어질 것이라고 믿습니다.

아프리카 사바나 코끼리의 의사소통 패턴에 대한 머신러닝 연구에 따르면, 코끼리들은 서로에게 이름을 지어주기도 합니다.

Way back in the 80s, I noticed that sometimes, when an elephant called a member of her family, one individual would answer, and everybody else ignored the calling animal. And then she would call again, and a different elephant would sort of lift her head up and rumble very loudly. This is Joyce Poole. She's been studying African elephants and their communication for 50 years. Then I started to think, well, okay, so maybe they have a way of directing a call to a specific individual. But we had no way of detecting that. Decades later, she partnered up with Mickey Pardo, who designed a study around her observations. I went out into the field. I recorded calls with careful behavioral observations. So we knew who made each call, we knew who the call was addressed to, we knew the context of the call... They encoded the acoustic information from the recordings into a long string of numbers, along with the data Mickey collected about the calls.

They fed nearly 500 different calls like this into a statistical model. And when given the acoustic structure of a new call, the model could predict who the receiver of the call was, much better than chance. In other words, evidence suggesting African savanna elephants give each other names. When we posted it on Facebook, somebody wrote back, said that the Earth just shifted a little bit. And I think that's true. This is just one example of how machine learning is decoding complexities in animal communication that humans can't detect. And now some AI researchers want to take the next step: Large language models, like the ones that power chatbots, but built for interspecies communication. “Can we talk a little bit about love?” “There is still much to be learned about whales.” When researchers study animal communication, they usually employ a few methods: Recording their vocalizations, observing and documenting the behavior and context around those sounds, and sometimes doing a playback to measure the animal's response.

All of these areas are already being impacted by AI. Recordings from the field don't usually sound like this. They often sound like this: Multiple animals vocalizing on top of one another in a noisy environment. This is known as the cocktail party problem, and it's a common issue in the field of animal research. But machine learning solved a similar problem in human speech recognition. AI researchers trained a model called Deep Karaoke on lots of music tracks where instruments and vocals were recorded separately, then also on the fully mixed tracks, until it was able to do the task of separating out instruments and vocals in new music clips. Recently, AI researchers have had some success applying similar algorithms to animal sound recordings. Which means you can take a clip of a group of macaque monkeys, and single out one discernable call. Researchers could also start using AI in how they use playbacks in the field. You may have seen AI models that can be trained on lots of examples of a sound recording

and then generate another unique version of it. AI researchers are starting to develop similar models for animal recordings. These are all types of “supervised learning.” That means that the model gets trained on lots of examples labeled by humans. And in the elephant name study, researchers were able to feed a model their observations, which, along with the sound data, helped them detect something in elephant calls they couldn't through observation alone. You need to annotate a lot of data. Yossi Yovel trained a statistical model on 15,000 Egyptian fruit bat vocalizations, and it was able to identify the emitter of the call, the context of the call, its behavioral response, and who the call was addressed to. And we annotated them manually. You know, I'm already saying this is a restriction of the study, because maybe we're missing something, we’re humans, we’re not bats. And that's the problem with supervised learning models. They are limited by what we humans already know about animal communication in order to label the training data.

And we don't know a lot. That's why some AI researchers say self-supervised models hold the most potential for decoding animal communication. This is how natural language processing models like ChatGPT are trained. Instead of human-labeled examples, they are trained on a large amount of unlabeled data, and then can sort it according to patterns and categories it detects all on its own. In the example of ChatGPT, it learned from all the books, websites, social media feeds, and anything else it could scrape from the internet, and came to its own conclusions about how language works. Every language has a shape that AI discovers. This is Aza Raskin. He co-founded the Earth Species Project, one of a few organizations that want to build models like this for animal communication. What he means by language having a shape is that language models are built out of relationships among words. Words that mean similar things are placed near each other, words that share a relationship share a distance and direction.

So, man is to king as woman is to queen. So this is the shape of all those relationships among the English language’s 10,000 most common words, visualized here by the Earth Species Project. Flattened out, it looks something like this. Something really miraculous happened in 2017, and that was researchers discovered that you could take the shape of any one language and match it to the shape of any other language, and the point which is “dog” ends up in the same spot. This idea, that similar words can be located in other languages in roughly the same place, is what gives the Earth Species Project hope we could do a version of this for animal communication. To do a translation without needing any examples, without needing a Rosetta stone. This is complicated though because we know that animals don't just communicate with sound, but with other senses, too. But Aza points out that we can learn from the fact that image generation models like DALL-E and Midjourney

are built on the same large language model structure used for text. It turns out, behind the scenes, it's again these kinds of shapes. There's the shape that represents sound, the shape that represents images. Those two shapes get aligned, and now you can translate between images and text. Their expectation is that where nonhuman animals’ communication would line up with ours will tell us even more about what we have in common. Dolphins look in mirrors and recognize themselves. Elephants too. That's a kind of self-awareness. One concern with this plan is related to a step in self-supervised learning called validation, meaning humans still need to refine these models by grading them on their answers. How would we do that in a communication so foreign from our own? We also might have too high expectations of this overlap, or the capacity for having a conversation with a nonhuman animal in a shared language, and about shared experiences. “Hey Kurt, how are you doing dude?”

“So I'm about to translate that into a meow.” “We said hi. Hi. Hi. Hi. You know, next time you want to say, how are you?” I do not think that humans should be considered more important than other species, but that doesn't mean that there's no usefulness in distinguishing between language, which is this very specific behavior that, at least based on what we currently know, seems to be unique to humans, and other forms of communication. In order to build these models, the first step is collecting a lot more data on animal sounds than exists right now. And so I'm actually at the moment building up a database with all the individual calls. So, close to 10,000 records in that, which is very small actually. Around the world, animal researchers are in the midst of a massive data collection effort, tagging and recording animals with video and sound and spatial data to feed these data-thirsty models. Time will tell whether true interspecies

communication will be facilitated by AI. But researchers hope that discoveries along the way will continue to have an impact on our appreciation and protection of the species we share the planet with. We're not the only ones on the planet who can communicate, who care about one another, who have thoughts about the past, and about the future. They also have a right to be here and a reason for being here.

How AI could help us talk to animals

Vox helps you understand our world.

Vox video is Bridgett Henwood, Cath Spangler, Adam Freelander, Joey Sendaydiego, Coleman Lowndes, Christophe Haubursin, Kim Mas, Laura Bult, Ed Vega, Rajaa Elidrissi, and with research support from Melissa Hirsch. 

Become a member today: https://www.vox.com/support-now

Check out our FAQ on permissions and licensing of our videos: https://www.voxmedia.com/pages/licensing

✉️ Write us: voxvideo@vox.com


아프리카 코끼리 의사소통 해독하기

조이스 풀은 수십 년 동안 아프리카 코끼리의 의사소통을 연구해 왔으며, 미키 파도의 도움으로 코끼리의 음향 구조를 바탕으로 코끼리가 전화 수신자를 예측할 수 있다는 증거를 발견했습니다.

이는 코끼리가 서로 이름을 부르는 것과 유사하게 특정 개인에게 호칭하는 방법이 있을 수 있음을 시사합니다.

현재 기계 학습은 동물 의사소통의 복잡성을 밝히는 데 사용되고 있으며, 종 간 의사소통을 위한 대규모 언어 모델을 개발하는 데 관심이 집중되고 있습니다.

동물 소리 연구에 대한 AI의 영향

AI는 사람의 음성 인식에서 해결책을 찾은 것과 마찬가지로 동물 소리 녹음에서 '칵테일 파티 문제'를 해결하는 데 성공했습니다.

연구자들은 이제 AI 알고리즘을 적용하여 녹음에서 다양한 동물 발성을 분리하고 분석하고 있습니다.

연구자들은 AI에서 지도 학습 모델을 사용하여 동물 울음소리에서 새로운 패턴과 정보를 감지할 수 있었지만, 인간 주석가의 기존 지식으로는 한계가 있었습니다.

동물 커뮤니케이션 해독을 위한 자기 감독 모델의 잠재력

ChatGPT와 같은 자가 지도 모델은 레이블이 없는 데이터를 학습하여 언어의 형태를 이해함으로써 동물의 의사소통을 해독합니다.

지구 종 프로젝트는 비슷한 단어가 여러 언어에 걸쳐 일관되게 배치된다는 점에 착안하여 동물의 의사소통을 위한 유사한 모델을 구축하는 것을 목표로 합니다.

소리 이상의 것을 포함하는 동물의 의사소통은 어려운 과제이지만, 연구자들은 이 모델이 다른 감각을 통해서도 학습할 수 있다고 믿습니다.

동물 커뮤니케이션을 이해하기 위한 언어 모델 사용

언어 모델은 인간과 동물의 의사소통 사이의 간극을 좁히고 공통점을 찾기 위해 사용되고 있습니다.

동물 의사소통의 이질적인 특성에도 불구하고 여전히 사람이 채점해야 하기 때문에 이러한 모델에서 자가 지도 학습의 검증에 대한 우려가 있습니다.

이 과정의 첫 번째 단계는 동물의 소리에 대한 더 많은 양의 데이터를 수집하는 것이며, 현재 전 세계 동물 연구자들이 대규모 데이터 수집 작업을 진행 중입니다.

AI가 종의 커뮤니케이션에 미치는 영향

인공지능은 다른 종과의 소통을 촉진하고 우리가 다른 종에 대한 이해와 보호에 기여할 것으로 기대됩니다.

다른 종들도 소통하고, 서로를 돌보고, 과거와 미래에 대해 생각할 수 있는 능력을 가지고 있어 지구상에서 자신의 존재와 존재의 중요성을 강조할 수 있습니다.