How AI could help us talk to animals

Back in the 1980s, Joyce Poole observed a fascinating behavior among African elephants. Whenever an elephant would call out to its family, only one member might respond, while others would ignore the call. This pattern intrigued her: was there a method behind these calls, possibly directing them to specific individuals? Despite her observations, there was no concrete way to verify this hypothesis at the time.

Decades later, Poole's curiosity led her to collaborate with Mickey Pardo, who designed a study based on her observations. Together, they ventured into the field, recording elephant calls with meticulous attention to behavioral contexts. By encoding these recordings into detailed numerical data and feeding nearly 500 different calls into a statistical model, they discovered something remarkable. The model could predict who the intended recipient of each call was, based on the acoustic structure, significantly better than random chance. This finding suggested that African savanna elephants might actually use a form of names in their communication.

When Poole and Pardo shared their findings on social media, the reaction was profound. One commenter even remarked that it felt as though "the Earth just shifted a little bit." This breakthrough is a prime example of how machine learning is revolutionizing our understanding of animal communication, uncovering complexities that were previously beyond human detection.

Currently, some researchers are exploring the next frontier: large language models designed for interspecies communication. These models, similar to those used in chatbots, aim to bridge the communication gap between humans and animals. Researchers typically use several methods to study animal communication: recording vocalizations, observing behavior and context, and using playback experiments to gauge responses. Each of these methods is being enhanced by advances in artificial intelligence (AI).

Field recordings of animal sounds often face challenges, particularly the "cocktail party problem," where multiple sounds overlap in a noisy environment. Machine learning has already addressed a similar issue in human speech recognition. For instance, the AI model Deep Karaoke was trained to separate vocals from music tracks, a technique that has been adapted to isolate individual animal calls from noisy recordings. This capability allows researchers to focus on specific vocalizations within a complex soundscape, such as identifying individual macaque monkey calls amid a cacophony of sounds.

AI is also improving playback experiments. Traditionally, researchers use playbacks to measure how animals react to specific sounds. AI models can now generate unique versions of sound recordings by learning from extensive examples, a process known as "supervised learning." This method relies on human-labeled data, which, while useful, is limited by our existing knowledge of animal communication. For instance, Yossi Yovel’s research on Egyptian fruit bats involved training a model on 15,000 vocalizations to identify who made each call and its context. However, the constraints of supervised learning mean that models can only be as accurate as the labeled data provided by humans, who may not fully understand the nuances of animal communication.

Self-supervised learning models, like those used in natural language processing for systems such as ChatGPT, offer a promising alternative. These models are trained on large amounts of unlabeled data, learning to detect patterns and categorize information independently. In ChatGPT’s case, the model analyzed vast amounts of text from the internet to understand language patterns. Aza Raskin, co-founder of the Earth Species Project, envisions a similar approach for animal communication. Just as language models map relationships among words, researchers hope to map the "shape" of animal communication to understand cross-species interactions.

Raskin’s Earth Species Project aims to create models that can interpret animal communication without needing pre-existing examples. This involves understanding how different forms of communication, such as sounds and images, can be translated between species. The project's goal is to uncover commonalities in communication, potentially allowing us to interpret animal languages in a way akin to translating between human languages.

One of the challenges is that animals communicate through multiple senses, not just sound. However, insights from image generation models like DALL-E and Midjourney, which use similar structures to language models, suggest that aligning different sensory data might reveal more about the overlaps between human and animal communication.

There are concerns about the validation process for these models. Self-supervised learning requires human input to refine and validate models, raising the question of how we can effectively assess communication systems so foreign to our own. Additionally, there is a risk of overestimating our ability to have meaningful conversations with nonhuman animals or imposing human-like expectations on their communication.

Despite these challenges, researchers are building extensive databases of animal sounds, collecting video, audio, and spatial data to feed into these AI models. This global effort aims to expand our understanding of animal communication and enhance our appreciation and protection of the species with whom we share the planet.

In conclusion, while AI has the potential to revolutionize our understanding of animal communication, translating this knowledge into meaningful interspecies dialogue remains a complex challenge. The ongoing efforts to decode animal languages not only promise to deepen our connection with other species but also highlight the intrinsic value of all forms of life on Earth.