Artificial and Biological Intelligence
Humans, Animals, and Machines
I believe a highly promising direction in AI research is to use artificial intelligence to better understand biological intelligence, and conversely, to use our understanding of biological intelligence to better understand how artificial intelligence works.
Understanding how AI works is the next frontier with a potential to bring enormous benefits to fields as diverse as scientific discovery, finding causal relationships in various data types, and introducing effective AI regulation. By studying human, animal, and artificial intelligence, we can enrich our understanding of the three components and build the next generation of AI models.
Understanding human language.
One of the main goals of the Berkeley Speech and Computation Lab, which I lead, is to develop artificial intelligence systems that learn in a way more similar to human infants.
Unlike large language models such as GPT-4, which are trained on massive amounts of text data by predicting the next token in a sequence, our models learn from raw speech. They learn in a fully unsupervised manner using imitation and imaginative learning. We introduced embodiment into AI models of language: we’ve designed the models to control representations of the mouth and tongue, enabling them to produce sounds that closely mimic sounds that they hear as the input. In other words, our models learn by listening to sounds of language and by trying to imitate them using articulators such as mouth and tongue.
We have demonstrated that the models can not only learn words and other processes in language, but can also create new English words that they have never encountered before. We also argue that symbolic, algebraic-like rules emerge within these purely connectionist models and provide a framework for exploring emergence of language-like properties in artificial neural networks devoid of human-specific attributes.
Understanding AI.
The second goal of my research is to better understand learning mechanisms of AI models. Interpretability of AI is the next frontier in machine learning research. Most interpretability work has centered on vision, a domain substantially more complex than spoken language. I argue that spoken language is the ideal testing ground for AI interpretability.
First, spoken language is fundamentally simpler than the visual world. Second, humans have an innate “generative” tool to produce speech — the articulators — while no such generative ability exists in the visual domain. For example, we can’t directly depict our visual imagination, but we can readily produce new words, sentences, and sounds with our mouth, tongue, and other articulators.
The AI industry has so far mainly centered on enhancing performance, leaving interpretability somewhat overlooked. However, understanding how AI learns is not only pivotal for gaining novel causal insights from data but is also key in developing effective regulations.
To address this, we have introduced several techniques that facilitate a deeper exploration into the inner workings of the models. For instance, we have demonstrated that the models can learn symbolic-like, discrete, and causal representations. Using our technique, we can introspect how the networks process language internally.
Our lab was the pioneer in illustrating that artificial neural networks and the human brain share striking similarities in processing language sounds, even when examining the raw, untransformed signals. This discovery gives us crucial insights for our understanding of human language acquisition. The proposed computational models can also serve as a basis for testing hypotheses that are ethically impossible to explore using the human brain.
While current large language models like GPT-4 are not realistic models of human language learning, they are instrumental for testing the limits of artificial neural intelligence. For example, our research has demonstrated that explicit recursion and a metalinguistic awareness thereof — a feature absent in non-human communication systems — can emerge in transformers trained on text. We also show that linguistic formalism can serve as a tool to interpret LLM’s metacognitive performance, an approach we refer to as behavioral interpretability.
Understanding animals.
Leveraging AI interpretability techniques opens up new avenues to analyze data where ground truth is absent. We use AI models that emulate human language learning to study whale vocalizations. The unknown in whale vocalization is not only what their communication means but also how to approach something as unknown as their vocalizations. Our strategy involves training networks that simulate whale communication and that embed information within whale vocalizations. We then employ the proposed interpretability techniques to introspect what meaningful properties the model had learnt.
AI models hold an advantage of being less biased compared to human analysis, capable of recognizing any significant data properties. While they don’t offer conclusive answers presently, they help narrow down the hypothesis space, paving the way for insightful post-hoc analytical verifications.
Our initial findings are encouraging, revealing not only that our methodology successfully identifies rules and properties previously presumed to be significant but also discerns new potentially meaningful properties, such as spectral mean and variance, that were not assumed meaningful before.
Two related questions.
Collectively, this research sheds light on fundamental questions about human intelligence. Language is the defining property of humans. While many animals have individual properties of language, none possess the exact language that humans have, not even our closest relatives such as bonobos and chimpanzees. By employing artificial neural networks to model human language, we are testing what can emerge in non-biological neural networks with no language- or human-specific properties.
This line of work addresses two pivotal and related questions. First, it explores whether animals exhibit rudimentary forms of language, with humans having advanced further in language evolution due to superior intelligence, or if there is a unique property exclusive to humans that enables language. Secondly — in a parallel fashion, it questions if the existing neural architectures in artificial intelligence are sufficient for achieving human-level intelligence, or if discovering new properties is essential to create highly intelligent models.
Answering these questions is crucial to deepen our understanding of what it means to be human and to discern how we are similar to and different from animals and machines.
I recommend that to make a decent investigation of cetacean languages or communication systems, one needs to be more open to the possibility that some cetaceans species may have equivalent and even more complex thought communication systems than ours, although different in certain properties. I wish you much success.