Welcome to the 25th edition of the WatchTower! In this edition, we explore how the latest AI technologies can help us uncover the mysteries of animal communication and discuss the ethical dilemmas that generative adversarial networks (GANs) present to artists.

📰 Featured in This Edition

• AI Reading Group

• AI Industry Night

• Earth Species Project - Using AI to Decode Animal Communication

• The Ethical Dilemma of GANs and Hand-drawn Art

🗓 Upcoming Events

Earth Species Project - Using AI to Decode Animal Communication

Human language undeniably played a decisive role in determining our fate as a species; it gave us the ability to cooperate on complex levels, create and share fictions, and ultimately escape the food chain to go on and build civilisations like those we live in today. This generally leads us to believe that human communication and social experiences are vastly superior to those of other animals. But what if we’re wrong? Recent advances in AI are offering unprecedented opportunities to challenge this assumption by analysing and decoding the complex communication systems of other species.  

Traditionally, methods of studying animal communication rely heavily on human interpretation – a painstaking process of observation and attempting to infer meaning – but these approaches are limited by our cognitive biases and sensory constraints. AI, on the other hand, can rapidly process and search through large datasets of animal sounds, movements, and other signals, uncovering patterns far beyond our perceptual reach. 

Tools like ChatGPT have already succeeded in revealing hidden structures within human language, utilising deep learning and transformer architectures to transform raw data into representations of words and their relationships in high-dimensional vector space. To visualise this, think of it as placing words in a 3D space. With enough data, the model autonomously learns to position ‘Man’ relative to ‘Woman’ in the same way it places ‘King’ relative to ‘Queen’ - with the arrows connecting them being the same size and direction - demonstrating that it understands these relationships are equivalent.

Similarly, AI tools may help us uncover hidden structures and meanings in the natural world, potentially revealing a level of complexity in animal communication that rivals our own. 

Earth Species Project

In fact, this work has already been underway since 2018, when Aza Raskin, Britt Selvitelle, and Katie Zacarian founded the Earth Species Project (ESP), a non-profit organisation dedicated to using AI to decode non-human communication. ESP is undertaking a wide range of ambitious projects that could have far-reaching applications and reveal profound ethical considerations we don’t yet fully understand. Let’s take a closer look at a few. 

1. Self-Supervised Ethogram Discovery 

To understand an animal’s behaviour, scientists create an inventory of the types of actions it performs, known as an ethogram. With bio-loggers (animal-attached tags) becoming smaller and lighter, there has recently been an explosion of audio, kinetic, and other types of data being collected from wild animals. ESP are using self-supervised learning, a technique to automatically discover and label patterns in data without human annotation, to automate the construction of ethograms. For example, models will learn to associate certain patterns in data with activities such as feeding and resting, and over long timespans will be able to construct a complete map of an animal’s behavioural patterns. 

2. Vocal Signaling in Endangered Beluga Whales 

ESP is using machine learning to categorise unlabeled calls of an endangered population of beluga whales in Canada. The project aims to map the extent of the species’ vocal repertoire and to determine whether dialect patterns are present within the species across different locations. The researchers believe that building a deeper understanding of the social structure of these whales will be crucial in minimising human impacts on them. 

3. Inter-Species Phonetic Alphabet (ISPA) for Transcribing Animal Sounds 

Imagine being able to translate animal vocalisations as if they were a foreign language. The Inter-Species Phonetic Alphabet (ISPA) is an ambitious project that aims to do just that. By creating a standardized phonetic alphabet for animal sounds, ISPA could revolutionise bioacoustics, allowing us to apply the same AI models used in human language to decode the rich and complex communication systems of other species.  

4. Solving the Cocktail Party Problem 

A major challenge in bioacoustics is the difficulty of isolating individual voices or sound sources from noisy environments, often forcing researchers to discard large amounts of valuable data. ESP has found a solution using neural networks, and have posted some impressive demonstrations with dogs, monkeys, and dolphins on their website. Check them out for yourself: Earth Species Source Separation Demo. This breakthrough will significantly increase data utilisation and be crucial in studying social species that often vocalise simultaneously in groups. 

5. Benchmarks and Foundation Models  

In AI, benchmarks are standardised tests used to measure the performance of models. For example, a model like ChatGPT might be tested on its reasoning, math, or coding abilities, and receive a score based on how well it performs. 

ESP is developing benchmarks specifically for models that work with animal vocalisations and movement. These benchmarks will allow researchers to assess how effectively their models extract meaning from animal data. If a model obtains low scores on a benchmark, failing to draw accurate conclusions from the benchmark datasets, it will signal to researchers that further refinement is needed before the model can be reliably used to explain their data. 

Foundation models are large AI models trained on huge amounts of data, making them versatile enough to assist with a wide range of tasks (e.g., ChatGPT). These models can also be fine-tuned, meaning they can be further trained on specific datasets to improve their performance on more specialised tasks. 

ESP is also developing foundation models specifically for animal communication. Being trained on large amounts of animal data, these models would be able to take in a researcher’s raw data and output structured insights in various forms, such as labelled data, classifications, or other meaningful interpretations. With fine-tuning as a feature, these models will offer researchers the best of both worlds: a broad understanding across many species, with the ability to zero in on the unique communication patterns of a particular species when further trained on additional data. 

Looking Ahead

Through AI, we are beginning to understand the communication of many species in ways that were once thought impossible. As we decode these communication systems, we may not only learn more about the animals themselves but also gain a deeper insight into what it means to be human in a world full of diverse and complex life. 

_{Published by Jonas Macken, September 09 2024.}

The Ethical Dilemma of GANs and Hand-drawn Art

1. Introduction to GAN Networks

Generative Adversarial Networks (GANs) are a type of AI model designed to create new data, such as images, by learning from existing ones. They consist of two neural networks—one generates images (the generator) and the other evaluates them (the discriminator). This process enables GANs to create highly realistic art that can mimic different styles.

2. How GANs Work

GANs are trained on vast datasets of existing images, including artwork by human artists. The generator network produces new images by learning patterns from this data, while the discriminator evaluates whether the generated image is close enough to real artwork. Over time, the system improves, generating images that can look like human-created art, often indistinguishable from originals.

3. Why It’s Concerning for Traditional Artists

For hand-drawn artists, GANs present several ethical concerns. First, GANs use existing art as training data, often without consent, raising issues around intellectual property. Furthermore, AI-generated art is faster and cheaper to produce, making it difficult for traditional artists to compete in the market. This devalues the time, skill, and emotional investment human artists put into their work.

4. Steps Ahead

To address these challenges, stronger copyright protections are needed to safeguard human artists’ work from being used without permission. Additionally, fostering collaboration between AI tools and human creativity could lead to new artistic possibilities, allowing both traditional and AI-generated art to coexist in a fairer, more ethical landscape.

_{Published by Lucy Lu, September 09 2024.}

Sponsors

Sources

• Earth Species Project - Using AI to Decode Animal Communication

Closing Notes

We welcome any feedback / suggestions for future editions here or email us at [email protected].

Stay curious,