5th-year Ph.D. student · Johns Hopkins University · kgarci18 [at] jhu.edu
Building human-aligned AI through computational cognitive neuroscience.
I am a 5th-year NSF Fellow pursuing my Ph.D. in Computational Cognitive Science at Johns Hopkins University, where I study how humans and machines understand social interactions in the Computational Cognitive Neuroscience Lab with Professor Leyla Isik.
I measure how people perceive dynamic social scenes, benchmark AI models against human brain and behavior, and use that human data to align video and language models with human social perception. My work has been published at NeurIPS (oral) and ICLR, and featured in The Wall Street Journal.
Previously, I earned my B.S. at Stanford University, worked as a fintech data scientist for Fortune 500 companies, and completed an NIH Postbac Fellowship in affective neuroscience and machine learning at Northwestern University, mentored by Professor Robin Nusslock, Dr. Zachary Anderson, and Dr. Cassandra VanDunk.
Humans read social scenes in a glance; AI still struggles. I use large-scale behavior, fMRI, and deep learning to understand how people perceive dynamic social interactions — and then use that human data to build models that see the social world the way we do.
Tens of thousands of odd-one-out judgments reveal the latent dimensions behind social perception.
350+ image, video, and language models tested against human brain and behavior.
A compact behavioral signal steers video and language models toward human social understanding.
NeurIPS'26 Oral49k human judgments nearly triple video-model alignment.
Preprint '26Learned dimensions explain 2× more than theory-based features.
NeurIPS'25 WSGrounding frozen LLMs in visual social interactions.
VSS'25 TalkWhat drives human similarity in dynamic social videos?
ICLR'25350+ models vs. human brain and behavior.
VSS/CCN'24Hierarchical DNN–brain correspondence in the lateral stream.
A compact human signal — 49,484 odd-one-out judgments — reshapes video-model geometry: V-JEPA 2.1 nearly triples its alignment with human social perception, beats caption-based language models, and transfers zero-shot to abstract social interactions.
53,223 odd-one-out judgments train a sparse embedding whose dimensions combine relationships with actions (caregiving vs. affectionate play, conversation, cooking), explain twice the variance of theory-based annotations, and better predict visual cortex.
Socially grounded visual tokens inserted into a frozen LLM sharply reduce perplexity on social dialogue and draw high attention at socially salient moments.
Caption embeddings — how people describe a scene — best match human similarity judgments, followed by intimacy ratings; judgments track body- and object-selective cortex, not scene areas.
Benchmarking 350+ image, video, and language models on social videos reveals a major gap in matching human social vision, especially in the lateral visual stream.
300+ DNNs show a feed-forward hierarchy in the lateral stream — but fall well short of the noise ceiling there, unlike in the ventral stream.
Reward-task fMRI features do not reliably predict dimensional anxiety and depression symptoms under linear SVR.
Deep neural networks have become powerful models of human object recognition. Can they also explain how humans perceive people interacting? I will present several studies from my doctoral work to identify a gap in existing models, characterize human social visual representations, and test how behavioral supervision can improve model alignment. First, I benchmark over 350 image, video, and language models against human behavioral ratings and fMRI responses to natural social videos. I find that models that approach the noise ceiling in ventral visual regions leave substantial variance unexplained in the lateral pathway, including superior temporal sulcus. Video models improve predictions in several mid-level lateral regions, while language models given human captions best predict social ratings. These results reveal a challenge for models that aim to explain both the brain and behavior for social vision. Then, in a complementary study, to understand the underlying representations of social perception in humans, we collect approximately 50,000 judgments of 250 social videos and identify dimensions of human social structure that combine relationships and actions and are able to predict representational structure in the human brain and behavior. Finally, I introduce behavioral geometric supervision (BGS), which uses these human similarity judgments to guide the learning of video models by combining local triplet constraints with a global RSA objective, updating fewer than 2% of model parameters. Across all models fine-tuned with BGS, human alignment is improved; particularly in V-JEPA 2.1, this nearly triples the variance explained in held-out human similarity judgments, reaching 78% of the estimated noise ceiling and surpassing the strongest caption embedding baseline. The resulting representations capture behavioral variance beyond caption embeddings, improve the readout of social-affective attributes without explicit attribute supervision, and, surprisingly, are even able generalize to out-of-distribution abstract social interactions. Together, these findings identify a behavioral target and demonstrate an effective path toward models of human social vision.