Neuroengineering · Behavior · Computation
We study how brains learn complex vocal behavior, build AI that reveals the structure of animal communication, and engineer new interfaces to the nervous system.
Photo: Arturo Álvarez-Buylla01 / Song learningUniversity of Oregon
Knight Campus
What we investigate
01We use songbirds to uncover how biological neural networks learn and produce complex, precisely timed behavior.
Our experiments span vocal learning, neural circuit dynamics, electrophysiology, and cellular-scale imaging in singing birds.
02We build self-supervised models at the natural timescale of birdsong, allowing machines to discover, parse, and compare vocal units with very little labeled data.
SongMAE combines 5 ms representations with spatial masking to resolve individual syllables while retaining strong species-classification and detection capabilities.
03We build miniature, long-lived interfaces for recording and stimulating the nervous system.
Projects include direct laser-write 3D printing, thin-film ultramicroelectrodes, miniature microscopes, and chronic peripheral nerve interfaces.
Where we are going next
Two new programs extend the lab’s work into mouse cortex: how neural populations encode temporal sequences, and how adult cortical information processing changes when plasticity is restored.
Earlier work in zebra finches revealed a temporal-to-spatial transformation across the auditory hierarchy. Primary neurons responded synchronously to individual clicks, while secondary areas mapped different interval sequences onto distinct population vectors. Birds discriminated sequences built from 11–40 ms intervals, but performance broke down when those intervals were lengthened—linking the structure of a population response to the timescale of auditory perception.
In mice, we will ask whether auditory cortex performs the same computation: integrating millisecond-scale sound history and converting it into sequence-selective population states. The goal is to determine where these codes emerge, what temporal range they preserve, and whether the separation between population states predicts which sound sequences an animal can distinguish.
Recent work found that Calbindin (Calb1) levels in inhibitory neurons determine the extent of visual-cortical plasticity, providing evidence that reactivating critical-period gene programs can restore juvenile-like plasticity in adult circuits.
Using electrophysiological methods, we will study how plasticity-inducing factors such as Calb1 affect the way adult mouse cortex represents and processes sensory information.
Featured research · bioRxiv preprint
Most bioacoustic encoders inherit design assumptions from human speech. SongMAE instead matches the faster temporal structure of birdsong, producing representations that organize naturally around syllables and generalize to species excluded from training.
Read the preprint Preprint; not yet peer reviewed.
Featured research · bioRxiv preprint
Expert time-frequency annotations are costly to produce across species and recording conditions. We asked Qwen3.8-27B to mark vocalizations in Xeno-Canto spectrograms, then used those boxes to train a compact SongMAE detector.
The student generalized across held-out collections and outperformed supervised detectors trained on human annotations for time-frequency localization. YOLO models trained on the same teacher labels also matched or exceeded counterparts trained on human labels—pointing to a scalable route from foundation models to specialized bioacoustic tools.
Read the preprint Preprint; not yet peer reviewed.
SongMAE helps us see the structure of song. This work asks which neural circuits keep that structure moving.
Featured research · bioRxiv preprint
Canary song depends not only on producing the right syllables, but on knowing when to move to the next one. By pairing targeted neural-circuit lesions with high-throughput TweetyBERT analysis of thousands of songs, we found that disruptions involving medial Anterior Forebrain Pathway circuitry destabilized phrase timing. Affected birds developed sustained, prolonged, and variable repetition of particular syllables, revealing an ongoing role for basal-ganglia circuits in fluent learned behavior.
Read the preprint “Stuttering-like” describes prolonged syllable repetition before transition; it does not imply direct equivalence with human developmental stuttering. Preprint; not yet peer reviewed.
About the lab
The Gardner Lab studies birdsong vocal production to understand how biological neural networks learn and produce complex behavior. Our work ranges from machine-learning analysis of behavior and electrophysiology to cellular-scale imaging in singing birds.
Alongside this basic science, we develop new neural interfaces designed to form stable, long-term connections with the brain and peripheral nervous system.
A growing focus is designing AI around the natural temporal structure of animal communication, rather than importing assumptions from human speech.
Lab community
Principal Investigator
Associate professor of neuroengineering studying songbirds, neural circuit dynamics, and next-generation brain interfaces.
Postdoctoral Scholar
Taylor studies how inhibitory-circuit gene programs regulate critical-period plasticity. Her recent work identified Calbindin (Calb1) levels in visual-cortex inhibitory neurons as a determinant of plasticity, showing that reactivating critical-period-stage gene expression can restore juvenile-like plasticity in adult circuits. In the Gardner Lab, she will investigate mechanisms of plasticity regulation through chronic neural recordings in mice.
PhD Student · Institute of Neuroscience
George develops self-supervised AI for understanding the structure of birdsong. His work includes TweetyBERT, which learns to parse unlabeled canary song into notes, syllables, and phrases, and SongMAE, a large-scale bioacoustic encoder designed around the natural timescale of birdsong. He also uses vision-language models to generate scalable training labels for detecting bird vocalizations in field recordings.
PhD Student · Institute of Neuroscience
Rose studies how basal ganglia circuits control the timing and sequencing of learned vocal behavior. Her work combines targeted experiments in adult canaries with high-throughput TweetyBERT analysis, revealing how disruption involving the medial anterior forebrain pathway destabilizes phrase timing and produces stuttering-like syllable repetition.
PhD Student · Knight Campus Bioengineering
Peter is working on self-supervised models of birdsong with an emphasis on JEPA architectures. Previously, he worked on non-invasive neural interfaces and low-cost microscopes.
Undergraduate
Cole applies the lab’s TweetyBERT machine-learning model to label birdsong and investigate the higher-order structure of song syntax.
Recent alumni
PhD Alumna · Knight Campus Bioengineering
During her PhD at the University of Oregon’s Knight Campus, Diana developed microfabricated devices and sensors for neuroscience. Her work focused on sensor design and on creating more effective interfaces with the peripheral and central nervous systems.
Publications & press
Explore peer-reviewed research in neural computation, vocal learning, machine learning, and neural interface engineering.
View Google ScholarSoftware resources
Browse code and workflows developed in the lab for analyzing birdsong and connecting machine learning with behavior.
A canary-specific self-supervised transformer trained exclusively on American Singer canary recordings. It learns notes, syllables, and phrases without human labels for automated song annotation and analysis.
A multispecies masked autoencoder pretrained at high temporal resolution on Xeno-canto field recordings, with Atlantic canaries excluded from pretraining. It resolves fine-grained syllables while retaining species-classification and song-detection abilities.
Join the conversation
We welcome conversations with prospective students, postdocs, collaborators, and people building new tools for neuroscience.
timg@uoregon.edu 1505 Franklin Boulevard, Eugene, Oregon 97403