Embedded Files
I'm interested in how humans, animals, and machines communicate and how intelligence – biological and artificial – is manifested through language. My research focuses on:
Interpretable AI for language – Building “artificial baby” models that learn speech from raw audio and tracing their inner workings against brain data.
AI-driven discovery – Exploring how our AI interpretability techniques can guide new insights for science.
Cross-species linguistics – Decoding sperm whale codas to test which properties of language extend beyond humans.
Origins of language – Modeling evolutionary pathways from vocal imitation to compositional syntax.
Law & society – Translating our interpretability findings into policy debates on the legal status of non-human animals.
Understanding how AI learns
Understanding how AI learns
We treat language as a microscope into both human and machine cognition.
Using latent-space probing and causal interventions, and other interpretability tools (IEEE ICASSP 2022, IEEE/ACM TASLP), we map where and how a model represents sounds, words, and grammar.
These methods reveal the model’s inner circuitry—which layers capture phonemes, which capture basic syntax, and how information flows between them (arXiv).
Language, it turns out, opens a clear window into the hidden workings of AI.
Building artificial baby language learners
Building artificial baby language learners
CiwGAN/fiwGAN (Neural Networks 2021) learn to speak the way human babies do:
hear a few dozen words of raw English audio,
invent new words such as start and dust (zero-shot generalization), and
move a virtual tongue and lips (ICASSP 2023).
How our GANs differ from text-based LLMs:
learn from raw speech, not text
succeed with tiny datasets
train by imitation / imagination (“imagitation”) rather than next-token prediction
have traces of communicative intent
include an explicit articulatory module
are evaluated with classic linguistic “wug” tests
These models let us ask how communicative systems emerge—and inform cognitive and neuroscience debates on the nature and origin of human language.
Comparing the brain and AI
Comparing the brain and AI
Using the artificial-baby network, we aligned its raw hidden activations with the human frequency-following response (FFR)—an early brainstem signal for speech.
The layer-to-brain correlation is one of the strongest model-brain matches reported to date.
Paper: “Encoding of Speech in Convolutional Layers and the Brainstem” (Scientific Reports, 2023)
Media: Quanta Magazine — “Some Neural Networks Learn Language Like Humans”
More
More
We have found one of the most similar signals between artificial intelligence agents and the human brain reported thus far, by comparing them directly.
These AI agents were trained to learn spoken language in a manner akin to how humans learn to speak: by immersing them in the raw sounds of language without supervision.
The study, published in Scientific Reports, is the first to directly compare raw brainwaves and AI signals without performing any transformations.
This line of work helps us better understand how AI learns, as well as identify similarities and differences between humans and machines.
🔊The sound played to humans and machines: link
🔊How this sound sounds like in the brain: link
🔊How this sound sounds like in machines1: link
🔊How this sound sounds like in machines2: link
Analyzing large language models
Analyzing large language models
Historically, models struggled with tasks that require “reflecting” on language rather than just using it.
We show that o1 is the first large language model that can not only do language, but also analyze language metalinguistically as a linguist would
Recursion is one of the few properties of human language not found in animals.
Remarkably, “o1” handled multiple layers of recursion, even extending sentences like the Nietzsche example to: “The worldview that the prose that the philosopher Nietzsche admired wrote expressed was unprecedented.”
Our proposed "behavioral interpretability" helps us understand the inner workings of the models.
Quote from the paper: “It appears that recursive reasoning with metacognitive awareness evolved in humans first and that similar behavior can emerge in deep neural network architectures trained on human language. It remains to be seen if animal communication in the wild or language-trained animals can approximate this recursive performance.”
Talk at the Simons Institute (Workshop on LLMs): link
Forthcoming in IEEE Transactions on AI
One of the first papers that tests GPT's coding abilities: arXiv
Using Generative AI to decode whale communication
Using Generative AI to decode whale communication
Road‑map paper – “Towards Understanding the Communication in Sperm Whales,” iScience 2022
Societal & legal impact – “What If We Understood What Animals Are Saying?” Ecology Law Quarterly
AI‑powered decoding of whale communication not only advances marine biology but also feeds into policy debates on non‑human rights, ocean‑noise regulation, and the criteria for legal personhood—a theme we develop in our Ecology Law Quarterly article.
Paper: Approaching an Unknown Communication System by Latent‑Space Exploration and Causal Inference (preprint)
We train GANs to imitate whale codas and embed information in them.
Introduce CDEV – Causal Disentanglement with Extreme Values, an interpretability trick that identifies which acoustic features the network treats as meaningful.
→ Predicted that spectral properties may carry information in whale dialogue. This was later confirmed by the following paper:
Paper: Vowels and Diphthongs in Sperm Whales (preprint)
We discovered that sperm whales have analogues to human vowels. We describe two recurring spectral patterns—an “a‑vowel” and an “i‑vowel” as well as diphthongs
We argue whales actively exchange these vowel analogues in conversational turns, and describe the structure behind the vowels.
Paper: The Phonology of Sperm Whale Coda Vowels (preprint)
We argue that whale vocalizations not only resemble human vowels, but also behave like ones.
We found distributional patterns, intrinsic duration, length distinction, and coarticulation in codas. We argue that this makes sperm whale codas one of the most linguistically and phonologically complex vocalizations and the one that is closest to human language.
Unnatural phonology
Unnatural phonology
How do cognitive constraints and historical forces each shape the world’s sound patterns? My work combines diachronic data, learning experiments, and statistical modeling to tease the two forces apart.
I argue that phonology offers a unique test case for distinguishing historical from cognitive influences on human behavior. The Language paper identifies a process called catalysis that explains how learning factors directly influence typology
I developed a statistical model for deriving typology within the “historical bias” approach (Phonology)
I established the Minimal Sound Change Requirement and the Blurring Process (Journal of Linguistics 2018)
I apply the Blurring Process to final nasalization (Glossa) and intervocalic devoicing (Journal of Linguistics 2024)
Indo-European linguistics
Indo-European linguistics
In the paper on Vedic meter (JAOS), I argue for a new rule in the Rigveda. I show that this new rule that restores the lost v and y sounds turns several “irregular” Rigvedic verses into perfectly regular metres.
I propose a new explanation for the development -aḥ > -o in Sanskrit and Avestan (Transactions of the Philological Society).
I propose new explanation of the phonetics of independent svarita: WeCIEC Proceedings.
In a project on Vedic pitch accent system, I combine philological and comparative sources with acoustic analyses of present-day Vedic recitation to provide a more accurate reconstruction of the Vedic accent, one of the oldest known accent marking systems.
Page updated
Google Sites
Report abuse