Jonathan C. Kao
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Research Subjects
Jonathan C. Kao has a total of 117 co-authors. Here are the top co-authors by number of shared publications:
- Johannes Y. Lee
University of California, Los Angeles, Los Angeles, USA
5 shared publications - Krishna V. Shenoy
Department of Electrical Engineering, Stanford University, Stanford, USA
5 shared publications - Michael Kleinman
Department of Electrical and Computer Engineering, University of California, Los Angeles
5 shared publications - Avishek Adhikari
Department of Psychology, University of California, Los Angeles, Los Angeles, USA
4 shared publications - Fernando Mcv Reis
Department of Psychology, University of California
4 shared publications Meghmik Chakerian
Department of Psychology, University of California, Los Angeles
4 shared publications
Latest Publications
- Top Cited PapersOpen Access
- 27 April 2026
- Published by Springer Nature in Nature Communications
- Vol. 17 (1)
Spurious alignment between large language models and brains can emerge from non-robust methods and overlooked confounds
Emerging research seeks to draw neuroscientific insights from the neural predictivity of large language models (LLMs). However, as results rapidly proliferate, there is a growing need for large-scale assessments of their robustness. Here, we analyze a wide range of models and methodological approaches across three widely used neural datasets. We find that the use of shuffled train-test splits has contributed to findings that are influential but spurious. Furthermore, how activations are extracted from LLMs can bias results in favor of specific model classes. Lastly, we find that confounding variables, particularly positional signals and word rate, perform competitively with trained LLMs and fully account for the neural predictivity of untrained LLMs on these neural datasets. Although many studies in the field avoid these pitfalls, our results indicate that some apparent alignment between LLMs and brains has emerged from non-robust methods and overlooked confounds.
- 27 April 2026
- Published by Springer Nature in Nature Communications
- Vol. 17 (1)
Top Cited PapersOpen AccessSpurious alignment between large language models and brains can emerge from non-robust methods and overlooked confounds
Emerging research seeks to draw neuroscientific insights from the neural predictivity of large language models (LLMs). However, as results rapidly proliferate, there is a growing need for large-scale assessments of their robustness. Here, we analyze a wide range of models and methodological approaches across three widely used neural datasets. We find that the use of shuffled train-test splits has contributed to findings that are influential but spurious. Furthermore, how activations are extracted from LLMs can bias results in favor of specific model classes. Lastly, we find that confounding variables, particularly positional signals and word rate, perform competitively with trained LLMs and fully account for the neural predictivity of untrained LLMs on these neural datasets. Although many studies in the field avoid these pitfalls, our results indicate that some apparent alignment between LLMs and brains has emerged from non-robust methods and overlooked confounds.
- Open Access
- 16 March 2026
- Published by Association for Computing Machinery (ACM)
- p. 982-990
DiSCo: Diffusion Sequence Copilots for Shared Autonomy
- 16 March 2026
- Published by Association for Computing Machinery (ACM)
- p. 982-990
Open AccessDiSCo: Diffusion Sequence Copilots for Shared Autonomy
- Top Cited Papers
- 1 January 2026
- Published by American Association for the Advancement of Science (AAAS) in Science
- Vol. 391 (6780) , eadw8151
Neural basis of cooperative behavior in biological and artificial intelligence systems
Cooperation, the process through which individuals work together to achieve common goals, is fundamental to human and animal societies and increasingly critical in artificial intelligence (AI). In this study, we investigated cooperation in mice and AI systems, examining how they learn to actively coordinate their actions to obtain shared rewards. We identified key social behavioral strategies and decision-making processes in mice that facilitate successful cooperation. These processes are represented in the anterior cingulate cortex (ACC), and ACC activity causally contributes to cooperative behavior. We extended our findings to AI systems by training artificial agents in a similar cooperation task. The agents developed behavioral strategies and neural representations reminiscent of those observed in the biological brain, revealing parallels between cooperative behavior in biological and artificial systems.
- 1 January 2026
- Published by American Association for the Advancement of Science (AAAS) in Science
- Vol. 391 (6780) , eadw8151
Top Cited PapersNeural basis of cooperative behavior in biological and artificial intelligence systems
Cooperation, the process through which individuals work together to achieve common goals, is fundamental to human and animal societies and increasingly critical in artificial intelligence (AI). In this study, we investigated cooperation in mice and AI systems, examining how they learn to actively coordinate their actions to obtain shared rewards. We identified key social behavioral strategies and decision-making processes in mice that facilitate successful cooperation. These processes are represented in the anterior cingulate cortex (ACC), and ACC activity causally contributes to cooperative behavior. We extended our findings to AI systems by training artificial agents in a similar cooperation task. The agents developed behavioral strategies and neural representations reminiscent of those observed in the biological brain, revealing parallels between cooperative behavior in biological and artificial systems.
- Top Cited Papers
- 1 September 2025
- Published by Springer Nature in Nature Machine Intelligence
- Vol. 7 (9) , 1-14
Brain–computer interface control with artificial intelligence copilots
Motor brain–computer interfaces (BCIs) decode neural signals to help people with paralysis move and communicate. Even with important advances in the past two decades, BCIs face a key obstacle to clinical viability: BCI performance should strongly outweigh costs and risks. To significantly increase the BCI performance, we use shared autonomy, where artificial intelligence (AI) copilots collaborate with BCI users to achieve task goals. We demonstrate this AI-BCI in a non-invasive BCI system decoding electroencephalography signals. We first contribute a hybrid adaptive decoding approach using a convolutional neural network and ReFIT-like Kalman filter, enabling healthy users and a participant with paralysis to control computer cursors and robotic arms via decoded electroencephalography signals. We then design two AI copilots to aid BCI users in a cursor control task and a robotic arm pick-and-place task. We demonstrate AI-BCIs that enable a participant with paralysis to achieve 3.9-times-higher performance in target hit rate during cursor control and control a robotic arm to sequentially move random blocks to random locations, a task they could not do without an AI copilot. As AI copilots improve, BCIs designed with shared autonomy may achieve higher performance.
- 1 September 2025
- Published by Springer Nature in Nature Machine Intelligence
- Vol. 7 (9) , 1-14
Top Cited PapersBrain–computer interface control with artificial intelligence copilots
Motor brain–computer interfaces (BCIs) decode neural signals to help people with paralysis move and communicate. Even with important advances in the past two decades, BCIs face a key obstacle to clinical viability: BCI performance should strongly outweigh costs and risks. To significantly increase the BCI performance, we use shared autonomy, where artificial intelligence (AI) copilots collaborate with BCI users to achieve task goals. We demonstrate this AI-BCI in a non-invasive BCI system decoding electroencephalography signals. We first contribute a hybrid adaptive decoding approach using a convolutional neural network and ReFIT-like Kalman filter, enabling healthy users and a participant with paralysis to control computer cursors and robotic arms via decoded electroencephalography signals. We then design two AI copilots to aid BCI users in a cursor control task and a robotic arm pick-and-place task. We demonstrate AI-BCIs that enable a participant with paralysis to achieve 3.9-times-higher performance in target hit rate during cursor control and control a robotic arm to sequentially move random blocks to random locations, a task they could not do without an AI copilot. As AI copilots improve, BCIs designed with shared autonomy may achieve higher performance.
- Open Access
LowKeyEMG: Electromyographic typing with a reduced keyset
We introduce LowKeyEMG, a real-time human-computer interface that enables efficient text entry using only 7 gesture classes decoded from surface electromyography (sEMG). Prior work has attempted full-alphabet decoding from sEMG, but decoding large character sets remains unreliable, especially for individuals with motor impairments. Instead, LowKeyEMG reduces the English alphabet to 4 gesture keys, with 3 more for space and system interaction, to reliably translate simple one-handed gestures into text, leveraging the recurrent transformer-based language model RWKV for efficient computation. In real-time experiments, participants achieved average one-handed keyboardless typing speeds of 23.3 words per minute with LowKeyEMG, and improved gesture efficiency by 17% (relative to typed phrase length). When typing with only 7 keys, LowKeyEMG can achieve 98.2% top-3 word accuracy, demonstrating that this low-key typing paradigm can maintain practical communication rates. Our results have implications for assistive technologies and any interface where input bandwidth is constrained.
Open AccessLowKeyEMG: Electromyographic typing with a reduced keyset
We introduce LowKeyEMG, a real-time human-computer interface that enables efficient text entry using only 7 gesture classes decoded from surface electromyography (sEMG). Prior work has attempted full-alphabet decoding from sEMG, but decoding large character sets remains unreliable, especially for individuals with motor impairments. Instead, LowKeyEMG reduces the English alphabet to 4 gesture keys, with 3 more for space and system interaction, to reliably translate simple one-handed gestures into text, leveraging the recurrent transformer-based language model RWKV for efficient computation. In real-time experiments, participants achieved average one-handed keyboardless typing speeds of 23.3 words per minute with LowKeyEMG, and improved gesture efficiency by 17% (relative to typed phrase length). When typing with only 7 keys, LowKeyEMG can achieve 98.2% top-3 word accuracy, demonstrating that this low-key typing paradigm can maintain practical communication rates. Our results have implications for assistive technologies and any interface where input bandwidth is constrained.
- Open Access
- 8 July 2025
- Published by Springer Nature in Scientific Reports
- Vol. 15 (1) , 24406
Multiregional representations of intertemporal decision making in human single neurons
Understanding the neural mechanisms underlying delay discounting—the tendency to prefer smaller, immediate rewards over larger, delayed rewards—is critical for elucidating the etiology of impulsive decision-making, a hallmark of several psychiatric conditions including substance use and impulse control disorders. Here, we investigate single-neuron activity in the orbitofrontal cortex (OFC), hippocampus, and amygdala of nine human participants performing a delay discounting task. Intracranial recordings yielded a total of 193 single units (50 OFC, 68 amygdala, and 75 hippocampus) and reveal distinct neural correlates of decision-making, including representations of choice preferences and decision difficulty across all three regions. Analyses demonstrate preferential encoding of choice in the OFC. Additionally, we report that hippocampal activity reflects interindividual differences in discounting rates, with stronger representation observed in participants with slower temporal discounting. These findings provide novel insights into the multiregional neural computations underlying intertemporal decision-making and their relationship to impulsive behaviors.
- 8 July 2025
- Published by Springer Nature in Scientific Reports
- Vol. 15 (1) , 24406
Open AccessMultiregional representations of intertemporal decision making in human single neurons
Understanding the neural mechanisms underlying delay discounting—the tendency to prefer smaller, immediate rewards over larger, delayed rewards—is critical for elucidating the etiology of impulsive decision-making, a hallmark of several psychiatric conditions including substance use and impulse control disorders. Here, we investigate single-neuron activity in the orbitofrontal cortex (OFC), hippocampus, and amygdala of nine human participants performing a delay discounting task. Intracranial recordings yielded a total of 193 single units (50 OFC, 68 amygdala, and 75 hippocampus) and reveal distinct neural correlates of decision-making, including representations of choice preferences and decision difficulty across all three regions. Analyses demonstrate preferential encoding of choice in the OFC. Additionally, we report that hippocampal activity reflects interindividual differences in discounting rates, with stronger representation observed in participants with slower temporal discounting. These findings provide novel insights into the multiregional neural computations underlying intertemporal decision-making and their relationship to impulsive behaviors.
- 2 July 2025
- Published by Springer Nature in Nature
- p. 1-11
Inter-brain neural dynamics in biological and artificial intelligence systems
Social interaction can be regarded as a dynamic feedback loop between interacting individuals as they act and react to each other1,2. Here, to understand the neural basis of these interactions, we investigated inter-brain neural dynamics across individuals in both mice and artificial intelligence systems. By measuring activities of molecularly defined neurons in the dorsomedial prefrontal cortex of socially interacting mice, we find that the multi-dimensional neural space within each individual can be partitioned into two distinct subspaces—a shared neural subspace that represents shared neural dynamics across animals and a unique neural subspace that represents activity unique to each animal. Notably, compared with glutamatergic neurons, GABAergic (γ-aminobutyric acid-producing) neurons in the dorsomedial prefrontal cortex contain a considerably larger shared neural subspace, which arises from behaviours of both self and others. We extended this framework to artificial intelligence agents and observed that, as social interactions emerged, so too did shared neural dynamics between interacting agents. Importantly, selectively disrupting the neural components that contribute to shared neural dynamics substantially reduces the agents’ social actions. Our findings suggest that shared neural dynamics represent a fundamental and generalizable feature of interacting neural systems present in both biological and artificial agents and highlight the functional significance of shared neural dynamics in driving social interactions.
- 2 July 2025
- Published by Springer Nature in Nature
- p. 1-11
Inter-brain neural dynamics in biological and artificial intelligence systems
Social interaction can be regarded as a dynamic feedback loop between interacting individuals as they act and react to each other1,2. Here, to understand the neural basis of these interactions, we investigated inter-brain neural dynamics across individuals in both mice and artificial intelligence systems. By measuring activities of molecularly defined neurons in the dorsomedial prefrontal cortex of socially interacting mice, we find that the multi-dimensional neural space within each individual can be partitioned into two distinct subspaces—a shared neural subspace that represents shared neural dynamics across animals and a unique neural subspace that represents activity unique to each animal. Notably, compared with glutamatergic neurons, GABAergic (γ-aminobutyric acid-producing) neurons in the dorsomedial prefrontal cortex contain a considerably larger shared neural subspace, which arises from behaviours of both self and others. We extended this framework to artificial intelligence agents and observed that, as social interactions emerged, so too did shared neural dynamics between interacting agents. Importantly, selectively disrupting the neural components that contribute to shared neural dynamics substantially reduces the agents’ social actions. Our findings suggest that shared neural dynamics represent a fundamental and generalizable feature of interacting neural systems present in both biological and artificial agents and highlight the functional significance of shared neural dynamics in driving social interactions.
- Open Access
A reinforcement learning based software simulator for motor brain-computer interfaces
Intracortical motor brain-computer interfaces (BCIs) are expensive and time-consuming to design because accurate evaluation traditionally requires real-time experiments. In a BCI system, a user interacts with an imperfect decoder and continuously changes motor commands in response to unexpected decoded movements. This “closed-loop” nature of BCI leads to emergent interactions between the user and decoder that are challenging to model. The gold standard for BCI evaluation is therefore real-time experiments, which significantly limits the speed and community of BCI research. We present a new BCI simulator that enables researchers to accurately and quickly design BCIs for cursor control entirely in software. Our simulator replaces the BCI user with a deep reinforcement learning (RL) agent that interacts with a simulated BCI system and learns to optimally control it. We demonstrate that our simulator is accurate and versatile, reproducing the published results of three distinct types of BCI decoders: (1) a state-of-the-art linear decoder (FIT-KF), (2) a “two-stage” BCI decoder requiring closed-loop decoder adaptation (ReFIT-KF), and (3) a nonlinear recurrent neural network decoder (FORCE).
Open AccessA reinforcement learning based software simulator for motor brain-computer interfaces
Intracortical motor brain-computer interfaces (BCIs) are expensive and time-consuming to design because accurate evaluation traditionally requires real-time experiments. In a BCI system, a user interacts with an imperfect decoder and continuously changes motor commands in response to unexpected decoded movements. This “closed-loop” nature of BCI leads to emergent interactions between the user and decoder that are challenging to model. The gold standard for BCI evaluation is therefore real-time experiments, which significantly limits the speed and community of BCI research. We present a new BCI simulator that enables researchers to accurately and quickly design BCIs for cursor control entirely in software. Our simulator replaces the BCI user with a deep reinforcement learning (RL) agent that interacts with a simulated BCI system and learns to optimally control it. We demonstrate that our simulator is accurate and versatile, reproducing the published results of three distinct types of BCI decoders: (1) a state-of-the-art linear decoder (FIT-KF), (2) a “two-stage” BCI decoder requiring closed-loop decoder adaptation (ReFIT-KF), and (3) a nonlinear recurrent neural network decoder (FORCE).
- Open Access
Non-invasive brain-machine interface control with artificial intelligence copilots
Summary: Motor brain-machine interfaces (BMIs) decode neural signals to help people with paralysis move and communicate. Even with important advances in the last two decades, BMIs face key obstacles to clinical viability. Invasive BMIs achieve proficient cursor and robotic arm control but require neurosurgery, posing significant risk to patients. Non-invasive BMIs do not have neurosurgical risk, but achieve lower performance, sometimes being prohibitively frustrating to use and preventing widespread adoption. We take a step toward breaking this performance-risk tradeoff by building performant non-invasive BMIs. The critical limitation that bounds decoder performance in non-invasive BMIs is their poor neural signal-to-noise ratio. To overcome this, we contribute (1) a novel EEG decoding approach and (2) artificial intelligence (AI) copilots that infer task goals and aid action completion. We demonstrate that with this “AI-BMI,” in tandem with a new adaptive decoding approach using a convolutional neural network (CNN) and ReFIT-like Kalman filter (KF), healthy users and a paralyzed participant can autonomously and proficiently control computer cursors and robotic arms. Using an AI copilot improves goal acquisition speed by up to 4.3 × in the standard center-out 8 cursor control task and enables users to control a robotic arm to perform the sequential pick-and-place task, moving 4 randomly placed blocks to 4 randomly chosen locations. As AI copilots improve, this approach may result in clinically viable non-invasive AI-BMIs.
Open AccessNon-invasive brain-machine interface control with artificial intelligence copilots
Summary: Motor brain-machine interfaces (BMIs) decode neural signals to help people with paralysis move and communicate. Even with important advances in the last two decades, BMIs face key obstacles to clinical viability. Invasive BMIs achieve proficient cursor and robotic arm control but require neurosurgery, posing significant risk to patients. Non-invasive BMIs do not have neurosurgical risk, but achieve lower performance, sometimes being prohibitively frustrating to use and preventing widespread adoption. We take a step toward breaking this performance-risk tradeoff by building performant non-invasive BMIs. The critical limitation that bounds decoder performance in non-invasive BMIs is their poor neural signal-to-noise ratio. To overcome this, we contribute (1) a novel EEG decoding approach and (2) artificial intelligence (AI) copilots that infer task goals and aid action completion. We demonstrate that with this “AI-BMI,” in tandem with a new adaptive decoding approach using a convolutional neural network (CNN) and ReFIT-like Kalman filter (KF), healthy users and a paralyzed participant can autonomously and proficiently control computer cursors and robotic arms. Using an AI copilot improves goal acquisition speed by up to 4.3 × in the standard center-out 8 cursor control task and enables users to control a robotic arm to perform the sequential pick-and-place task, moving 4 randomly placed blocks to 4 randomly chosen locations. As AI copilots improve, this approach may result in clinically viable non-invasive AI-BMIs.
- Open Access
Gacs-Korner Common Information Variational Autoencoder
We propose a notion of common information that allows one to quantify and separate the information that is shared between two random variables from the information that is unique to each. Our notion of common information is defined by an optimization problem over a family of functions and recovers the G\'acs-K\"orner common information as a special case. Importantly, our notion can be approximated empirically using samples from the underlying data distribution. We then provide a method to partition and quantify the common and unique information using a simple modification of a traditional variational auto-encoder. Empirically, we demonstrate that our formulation allows us to learn semantically meaningful common and unique factors of variation even on high-dimensional data such as images and videos. Moreover, on datasets where ground-truth latent factors are known, we show that we can accurately quantify the common information between the random variables.
Open AccessGacs-Korner Common Information Variational Autoencoder
We propose a notion of common information that allows one to quantify and separate the information that is shared between two random variables from the information that is unique to each. Our notion of common information is defined by an optimization problem over a family of functions and recovers the G\'acs-K\"orner common information as a special case. Importantly, our notion can be approximated empirically using samples from the underlying data distribution. We then provide a method to partition and quantify the common and unique information using a simple modification of a traditional variational auto-encoder. Empirically, we demonstrate that our formulation allows us to learn semantically meaningful common and unique factors of variation even on high-dimensional data such as images and videos. Moreover, on datasets where ground-truth latent factors are known, we show that we can accurately quantify the common information between the random variables.