ChatGPT uses its own way to understand the world. Can similar technology be used to learn the language of animals? Solomon was able to communicate with animals not because he possessed magic items, but because of his gift of observation. ——Conrad Lorenz, "King Solomon's Ring" In animal-centered works such as "The Lion King" and "Zootopia", the author often personifies the characters and uses human thinking and communication methods to advance the plot.
However, this type of work can also lead to cognitive dissonance. When we communicate with animals, we may project our own ideas and prejudices onto the animals. For example, "the lamb kneeling to breastfeed" has nothing to do with gratitude or filial piety, but because of the special stomach structure of the sheep, but humans will project themselves onto the behavior of the lambs.
Traditional animal cognition work mainly focuses on establishing a vocabulary, but concepts such as "water", "drink", and "dry" may not exist or have meaning in the world of aquatic organisms, and there is no correspondence with human concepts in animal communication; and communication between animals does not necessarily involve vocalizations, but also includes gestures, action sequences, or changes in skin texture.
Theoretically, machine learning models are better than humans at summarizing loose correlations between words. The input to the neural network does not make any assumptions about the nature of the input data. As long as a certain pattern appears frequently, it is possible to discover the information contained in animal communication.
The Cetacean Translation Initiative (CETI), launched by research institutions such as City University of New York, UC Berkeley, MIT, Harvard, Google Research, and National Geographic, uses natural language processing systems to analyze massive sperm whale data and plans to directly communicate with wild sperm whales in the future.
The Earth Species Project (ESP) co-founded by Aza Raskin and others has open sourced the first animal vocalization benchmark BEANS, which can measure the performance of machine learning algorithms on bioacoustic data; it has also developed the first basic model for animal vocalizations AVES, which can be used for various tasks such as signal detection and classification.
With the advancement of generative AI technology, it may one day be possible to uncover the true meaning behind animal communication.
The complex animal kingdom
In 1974, the philosopher Thomas Nagel published a seminal paper called "What's it Like to Be a Bat?" "(WhatIsItLiketoBeaBat?"), he believes that the life of a bat is so different from that of humans that humans may never truly know the answer to this question.
Our understanding of the world is shaped by human concepts, and the only way to know what a bat is like is to be a bat and have the concept of a bat.
However, we can still infer part of the way bats think. For example, bats live at high places, and the concept of up and down may be reversed, through echolocation, etc., but we cannot have the life experience of bats.
If lions could talk, we wouldn't be able to understand it because the human brain cannot empathize with the feelings and concepts conveyed in lion language. ——Ludwig Wittgenstein
But not all animals think very differently from humans. Psychologically speaking, humans have more in common with other primates than octopuses and squids: Our last common ancestor with chimpanzees lived 6 million to 8 million years ago, and our last common ancestor with octopuses lived in the Precambrian oceans about 600 million years ago.
After being taught, chimpanzees can learn human sign language, and can even understand complex human instructions and communicate using keyboard symbols, but as mentioned at the beginning, we may also be over-anthropomorphizing the behavior of chimpanzees.
For species more distantly related to humans, understanding how they communicate becomes more difficult. For example, bees and some birds can see ultraviolet light in the visible spectrum, and bats, dolphins, dogs and cats can hear ultrasonic waves. Each species has its own unique characteristics.
Using AI to understand animals
Britt Selvitelle, a computer scientist at the Earth Species Project, said they are working on deciphering the first non-human language and may be able to do so within five to 10 years.
In the field of animal language, although researchers have accumulated a lot of knowledge for decades, there is still no "Rosetta Stone" in the world that can translate human language and animal language, and there is no gold standard for labeling "animal language."
Fundamentally, artificial intelligence is a data-driven tool. Pre-trained language models can learn the internal representation of data in an unsupervised manner through massive data.
Judging from the powerful performance of ChatGPT, generative AI technology may have its own unique internal representation method instead of applying human concepts, so researchers began to turn to AI technology to analyze data and obtain meaningful terms for animals.
In the Earth Species Project, data is collected in the form of sound, movement and video, covering animals in the wild or in captivity, and is accompanied by biologists' notes on what the animals were doing at the time and in what context.
As the Internet of Things matures, it is becoming easier to put cheap and reliable recording devices (such as microphones or biometric recorders) on animals in the wild, which can provide large amounts of data for artificial intelligence tools to organize and analyze to help discover the meaning behind the data, and then use generative methods to test, and ultimately achieve re-creation of animal sounds for two-way communication.
ANIMAL SOUNDS BEANS
In the field of bioacoustics, the successful application of machine learning-based techniques requires a carefully curated set of high-quality data on a specific task, but until now there has not been a public benchmark covering multiple tasks and multiple species to measure the performance of machine learning techniques in a controlled and standardized way and to benchmark newly proposed techniques against existing techniques.
Paper link: https://arxiv.org/pdf/2210.12300.pdf
Data link: https://github.com/earthspecies/beans
BEANS (theBEnchmark of ANimalSounds, the Benchmark of Animal Sounds) is a collection of bioacoustic tasks and public datasets specifically designed to measure the performance of machine learning algorithms in the field of bioacoustics, including two common tasks in bioacoustics: classification and detection.
BEANS includes 12 datasets covering multiple species, including birds, terrestrial and marine mammals, anurans and insects.
In addition to the dataset, the paper also proposes the performance of a set of standard machine learning methods as a baseline for task performance.
Both the benchmark and the baseline code are open source and the researchers hope that BEANS can establish a new standard data set for machine learning-based bioacoustic research.
Animal vocalization model AVES
In the field of bioacoustics, the lack of annotated training data has greatly hindered the use of large-scale neural network models trained in a supervised manner in this field.
In order to utilize a large amount of unlabeled audio data, researchers proposed AVES (Animal Vocalization Encoder based on Self-Supervision, self-supervised animal vocalization encoder), a self-supervised, Transformer model-based audio representation model that can be used to encode animal vocalizations.
Paper link: https://arxiv.org/pdf/2210.14493.pdf
Model link: https://github.com/earthspecies/aves
The researchers pre-trained the AVES model on a different set of unlabeled audio datasets and fine-tuned the model for downstream bioacoustic tasks.
Comprehensive experiments on classification and detection tasks show that AVES outperforms all strong baselines and even outperforms supervised topline models trained on annotated audio classification datasets.
Experimental results also show that carefully designing a small training subset relevant to downstream tasks is an effective way to train high-quality audio representation models.
ethical issues
In the 1970s, when Western society first discovered the song of whales, human society suspended the hunting of deep-sea whales and led to the establishment of the Environmental Protection Agency.
As the Earth Species Project technology roadmap advances, we can learn more about the organisms around us, collect more data, and develop new baselines and foundational models so we can better protect our blue planet.
Raskin believes that within the next 12-36 months, the team will be able to communicate with animals, such as making an artificial whale or crow that can talk to the whale or crow in an indistinguishable way, but the key point is that we also need to understand what the model is saying before we can further the conversation.
The Raskin team is also discussing how to use these artificial intelligence methods responsibly. It has now been stipulated that these methods must be prepared in any test. The technical roadmap points out potential risks, such as interfering with hunting and foraging or mating, and may also send errors to animals.
Humans only learned how to speak and communicate using sounds between 100,000 and 300,000 years ago, while whales and dolphins have been using sounds to pass down culture and songs for 34 million years.
If AI audio is randomly sent among whales, it may cause damage to 34 million years of culture.
That’s why much of the work so far in the Planet Species Project has been in collecting data and creating the foundations—the baselines and foundational models that will drive future progress—not unlike what companies and organizations around the world are doing every day with artificial intelligence and machine learning, just on a much grander scale.
If AI can help us understand what animals are saying, what are the limits to our ability to use AI?
If AI can help us understand animals, what will it teach us about humans?
Raskin and Zacarian hope the eventual translation of animal language becomes one of those turning points in world history, moments like the first discovery of a whale's song or the 1990 photograph of APale Blue Dot, which changed our view and understanding of the world.