Catalin Mitelut, JD, PhD
Neuroscience and AI researcher
I am a postdoctoral researcher at University of Basel and NYU working on the neuroscience of agency in biological organisms and artificial neural networks. My main neuroscience research involves developing behavior paradigms for understanding causality in free, volitional action - while in parallel characterizing the limits of behavior prediction and future pathways for human behavior manipulation from AI systems.
Background
Research
(select publications)
Agency in biological organisms
-
We show that neural correlates of self-initiated action are similar between mice and humans and that upcoming behaviors could be decoded seconds prior to movement.
Mitelut et al (2023). "The emergence of autonomous behavior in freely behaving gerbil pups. (in progress)
-
Using machine vision tools we track the behavior of individual gerbil pups across development and identify distinct developmental trajectories for social and non-social behaviors.
Mitelut et al (2023). "OpenCaBMI: an open source tool for two-photon imaging brain-machine-interface protocols
-
We developed a python-based package that enables the implementation of learning-based BMI paradigms and we show the acquisition of single neuron learning in mice.
Mitelut et al (2023). "Decoding volitional intent from the mouse hippocampus" (in progress)
-
We developed methods for decoding future intentions of mice from real-time neural activity.
Mitelut et al (2023) "Agency hackathon"
-
A hackathon for advancing our understanding of agency in artificial intelligence research by focusing on RL/IRL, mechanistic interpretability, game theory and other concepts.
-
We show that intent-aligned AI systems pose a risk to human agency (i.e. control over the world) and suggest several research paradigms aimed at helping protect human agency in human-AI interactions.
Agency in artificial intelligence models
Mitelut C (2022), "LLMs may capture key components of human agency" (Lesswrong)
-
I briefly discuss the neuroscience and psychology of agency in humans and the possible connection to large-language-models (LLMs) tendencies to confidently confabulate - which is a common phenomenon in human experience of agency.
Machine learning and statistics
-
We used statistics and machine learning to develop state of the art spike-sorting algorithms that significantly outperform other methods.
-
We used simulated electrically active neural networks to train and tune spike sorting algorithms.
Modeling of biological neural networks
Mitelut C et al (2015), "Standardizing spike sorting: an in vitro, in silico and In vivo study to develop quantitative metrics for sorting extracellularly recorded spiking activity", SFN 2015.
-
Using super-computer clusters we developed the first dataset of extracellular potentials from models of mouse V1
-
A large-scale project to model the mouse V1 at the Allen Institute for Brain Science. I contributed to the extracellular potential pipelines and OpenGL visualization code. My visualization package made the cover of Neuron (from Billeh et al 2020).
-
We used simulated electrically active neural networks to test electrode configurations for the Neuropixels probes.
Neural data science
-
We performed simultaneous widefield calcium and extracellular electrophysiology recordings to link single spiking neuron activity to cortex-wide networks.
