John Pearson
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Previous Organizations / Institutes
The organizations / institutes from the scholar's other articles in the past ten years
Research Subjects
John Pearson has a total of 176 co-authors. Here are the top co-authors by number of shared publications:
Michael L. Platt
Center for Cognitive Neuroscience and Duke Institute for Brain Sciences, Duke University, Durham, NC, United States
12 shared publications- Richard Mooney
Department of Neurobiology, Duke University
11 shared publications - Benjamin Yost Hayden
Baylor College of Medicine
7 shared publications - Greg D. Field
Stein Eye Institute, Department of Ophthalmology, University of California, Los Angeles, USA
5 shared publications - Na Young Jun
Department of Neurobiology, Duke University
5 shared publications - Scott A. Huettel
Duke University
5 shared publications
Latest Publications
- Open Access
- 7 September 2026
- Published by eLife Sciences Publications, Ltd in eLife preprints
- Vol. 15
Correctness is its own reward: bootstrapping error signals in self-guided reinforcement learning
Reinforcement learning (RL) offers a compelling account of how agents learn complex behaviors by trial and error, yet RL is predicated on the existence of a reward function provided by the agent’s environment. By contrast, many skills are learned without external guidance, posing a challenge to RL’s ability to account for self-directed learning. For instance, juvenile male zebra finches first memorize and then train themselves to reproduce the song of an adult male tutor through extensive practice. This process is believed to be guided by an internally computed assessment of performance quality, though the mechanism and development of this signal remain unknown. Here, we propose that, contrary to prevailing assumptions, tutor song memorization and performance assessment are subserved by the same neural circuit, one trained to predictively cancel tutor song. To test this hypothesis, we built models of a local forebrain circuit that uses contextual premotor signals to cancel tutor song auditory input via synaptic plasticity. After learning, excitatory projection neurons signaled mismatches between the tutor song and birds’ own performance, best matching experimental data when learning involved anti-Hebbian plasticity in recurrent interneurons. We also found that learning proceeds by both sharpening error sensitivity and minimizing responses to the tutor song. Finally, the error signals produced by this model can train a simple RL agent to replicate the spectrograms of adult bird songs. These results suggest that local learning via predictive cancellation suffices for bootstrapping error signals capable of guiding self-directed learning of natural behaviors.
- 7 September 2026
- Published by eLife Sciences Publications, Ltd in eLife preprints
- Vol. 15
Open AccessCorrectness is its own reward: bootstrapping error signals in self-guided reinforcement learning
Reinforcement learning (RL) offers a compelling account of how agents learn complex behaviors by trial and error, yet RL is predicated on the existence of a reward function provided by the agent’s environment. By contrast, many skills are learned without external guidance, posing a challenge to RL’s ability to account for self-directed learning. For instance, juvenile male zebra finches first memorize and then train themselves to reproduce the song of an adult male tutor through extensive practice. This process is believed to be guided by an internally computed assessment of performance quality, though the mechanism and development of this signal remain unknown. Here, we propose that, contrary to prevailing assumptions, tutor song memorization and performance assessment are subserved by the same neural circuit, one trained to predictively cancel tutor song. To test this hypothesis, we built models of a local forebrain circuit that uses contextual premotor signals to cancel tutor song auditory input via synaptic plasticity. After learning, excitatory projection neurons signaled mismatches between the tutor song and birds’ own performance, best matching experimental data when learning involved anti-Hebbian plasticity in recurrent interneurons. We also found that learning proceeds by both sharpening error sensitivity and minimizing responses to the tutor song. Finally, the error signals produced by this model can train a simple RL agent to replicate the spectrograms of adult bird songs. These results suggest that local learning via predictive cancellation suffices for bootstrapping error signals capable of guiding self-directed learning of natural behaviors.
- 19 May 2026
- Published by Society for Neuroscience in Journal of Neuroscience
- Vol. 46 (25)
Selective Encoding of Priors for Flexible Categorization But Not Bayesian Inference in the Frontal Eye Field
Expectations or prior beliefs about the world modulate sensory processing at the behavioral and neural levels. Bayesian models predict that such priors compensate for input uncertainty to optimize sensory judgments. Although Bayesian behavior is prevalent across sensorimotor systems, the relationship between priors and Bayesian inference is not obligatory. Priors may simply shift one’s internal decision boundaries without interacting with sensory uncertainty at all. We recently showed that humans and monkeys use both Bayesian and non-Bayesian strategies when reporting judgments of visual stability across saccades, despite using priors in both cases. While they increased prior use to compensate for internal, movement-driven sensory uncertainty in a Bayesian manner, they decreased prior use when faced with external, visual image uncertainty. The latter, “anti-Bayesian” pattern was best explained by a model in which category boundaries were adjusted by the prior but susceptible to image noise. Here, in two male rhesus macaques we recorded neural activity in the frontal eye field (FEF), a prefrontal region important for visuosaccadic behavior. We toggled between subjects’ prior use for Bayesian and anti-Bayesian behavior via trial-by-trial manipulation of the two uncertainty conditions. First, we found that FEF activity signaled the priors in both conditions. The prior-related modulation of activity, however, predicted only the anti-Bayesian, categorization behavior. The results suggest that neural activity in the FEF reflects the use of a flexible decision boundary for the perception of visual stability and, more generally, that neural mechanisms for Bayesian inference and visual categorization are dissociable and distributed in the primate brain. Significance Statement Appreciating a visual scene depends not only on retinal input, but also on priors about the world. Foreknowledge interacts with visual inputs to improve reactions and decisions. One way the brain combines priors and inputs is by using Bayes’ rule to model optimal outcomes. A simpler way is by categorizing inputs with prior-adjusted boundaries. Here, we tested how neurons in primate frontal cortex use priors: for Bayes’ rule, or for flexible categorization? A key feature of the study was to use a single perceptual task that was varied trial-by-trial to yield either Bayesian or categorization behaviors. We could then establish which behavior the neurons encoded. The implications extend beyond visuomotor behavior to broader neurocomputational mechanisms of prior use for cognition.
- 19 May 2026
- Published by Society for Neuroscience in Journal of Neuroscience
- Vol. 46 (25)
Selective Encoding of Priors for Flexible Categorization But Not Bayesian Inference in the Frontal Eye Field
Expectations or prior beliefs about the world modulate sensory processing at the behavioral and neural levels. Bayesian models predict that such priors compensate for input uncertainty to optimize sensory judgments. Although Bayesian behavior is prevalent across sensorimotor systems, the relationship between priors and Bayesian inference is not obligatory. Priors may simply shift one’s internal decision boundaries without interacting with sensory uncertainty at all. We recently showed that humans and monkeys use both Bayesian and non-Bayesian strategies when reporting judgments of visual stability across saccades, despite using priors in both cases. While they increased prior use to compensate for internal, movement-driven sensory uncertainty in a Bayesian manner, they decreased prior use when faced with external, visual image uncertainty. The latter, “anti-Bayesian” pattern was best explained by a model in which category boundaries were adjusted by the prior but susceptible to image noise. Here, in two male rhesus macaques we recorded neural activity in the frontal eye field (FEF), a prefrontal region important for visuosaccadic behavior. We toggled between subjects’ prior use for Bayesian and anti-Bayesian behavior via trial-by-trial manipulation of the two uncertainty conditions. First, we found that FEF activity signaled the priors in both conditions. The prior-related modulation of activity, however, predicted only the anti-Bayesian, categorization behavior. The results suggest that neural activity in the FEF reflects the use of a flexible decision boundary for the perception of visual stability and, more generally, that neural mechanisms for Bayesian inference and visual categorization are dissociable and distributed in the primate brain. Significance Statement Appreciating a visual scene depends not only on retinal input, but also on priors about the world. Foreknowledge interacts with visual inputs to improve reactions and decisions. One way the brain combines priors and inputs is by using Bayes’ rule to model optimal outcomes. A simpler way is by categorizing inputs with prior-adjusted boundaries. Here, we tested how neurons in primate frontal cortex use priors: for Bayes’ rule, or for flexible categorization? A key feature of the study was to use a single perceptual task that was varied trial-by-trial to yield either Bayesian or categorization behaviors. We could then establish which behavior the neurons encoded. The implications extend beyond visuomotor behavior to broader neurocomputational mechanisms of prior use for cognition.
- 13 May 2026
- Published by Springer Nature in Nature
- Vol. 655 (8124) , 1-9
A synaptic locus of song learning
Learning by imitation is the foundation for verbal and musical expression, but its neural basis remains unclear. A juvenile male zebra finch imitates the multisyllabic song of an adult tutor in a process that depends on a song-specialized cortico-basal ganglia circuit1,2,3,4, affording a powerful system to identify the synaptic substrates of imitative motor learning. Plasticity at a particular set of cortico-basal ganglia synapses is hypothesized to drive rapid learning-related changes in song before these changes are subsequently consolidated in downstream circuits5. Nevertheless, this hypothesis is untested and the synaptic locus where learning initially occurs is unclear. Here, by combining a computational framework to quantify song learning with synapse-specific optogenetic and chemogenetic manipulations within and downstream of the cortico-basal ganglia circuit, we identified the specific cortico-basal ganglia synapses that drive the acquisition and expression of rapid vocal changes during juvenile song learning and characterized the hours-long timescale over which these changes consolidate. Furthermore, transiently augmenting postsynaptic activity in the basal ganglia briefly accelerates learning rates and persistently alters song, demonstrating a direct link between basal ganglia activity and rapid learning. These results localize the specific cortico-basal ganglia synapses that enable a juvenile songbird to learn to sing and reveal the circuit logic and behavioural timescales of this imitative learning paradigm.
- 13 May 2026
- Published by Springer Nature in Nature
- Vol. 655 (8124) , 1-9
A synaptic locus of song learning
Learning by imitation is the foundation for verbal and musical expression, but its neural basis remains unclear. A juvenile male zebra finch imitates the multisyllabic song of an adult tutor in a process that depends on a song-specialized cortico-basal ganglia circuit1,2,3,4, affording a powerful system to identify the synaptic substrates of imitative motor learning. Plasticity at a particular set of cortico-basal ganglia synapses is hypothesized to drive rapid learning-related changes in song before these changes are subsequently consolidated in downstream circuits5. Nevertheless, this hypothesis is untested and the synaptic locus where learning initially occurs is unclear. Here, by combining a computational framework to quantify song learning with synapse-specific optogenetic and chemogenetic manipulations within and downstream of the cortico-basal ganglia circuit, we identified the specific cortico-basal ganglia synapses that drive the acquisition and expression of rapid vocal changes during juvenile song learning and characterized the hours-long timescale over which these changes consolidate. Furthermore, transiently augmenting postsynaptic activity in the basal ganglia briefly accelerates learning rates and persistently alters song, demonstrating a direct link between basal ganglia activity and rapid learning. These results localize the specific cortico-basal ganglia synapses that enable a juvenile songbird to learn to sing and reveal the circuit logic and behavioural timescales of this imitative learning paradigm.
- Open Access
Dynamic Compression Flows for Neuroscience Data
While neuroscience experiments have repeatedly demonstrated the involvement of large populations of neurons in even simple behaviors, these studies have just as often reported that the collective dynamics of neural activity are approximately low-dimensional. As a result, methods for identifying low-dimensional latent representations of time series data have become increasingly prominent in neuroscience. However, most existing methods either ignore temporal structure or model time evolution using latent dynamical systems approaches. In the first case, dynamics may be distorted or even scrambled in the latent space, while in the second, many possible latent dynamics may give rise to the same data. Here, we address these challenges using a novel flow-matching approach in which data are generated by a pair of flow fields, one governing time evolution, the other a mapping between data and a low-dimensional latent space. Importantly, the dimension-reducing flow is trained to minimize distortions of the temporal dynamics, learning an identifiable low-dimensional representation that preserves temporal relations in the original data. Additionally, we constrain our latent spaces to have low-dimensional support in a soft, parameterized manner, taking inspiration from ideas on nested dropout. Across both neural and behavioral data, we show that this dual flow approach produces both more interpretable dynamics and higher-quality reconstructions than competing models, including in noise-dominated data sets where conventional approaches fail.
Open AccessDynamic Compression Flows for Neuroscience Data
While neuroscience experiments have repeatedly demonstrated the involvement of large populations of neurons in even simple behaviors, these studies have just as often reported that the collective dynamics of neural activity are approximately low-dimensional. As a result, methods for identifying low-dimensional latent representations of time series data have become increasingly prominent in neuroscience. However, most existing methods either ignore temporal structure or model time evolution using latent dynamical systems approaches. In the first case, dynamics may be distorted or even scrambled in the latent space, while in the second, many possible latent dynamics may give rise to the same data. Here, we address these challenges using a novel flow-matching approach in which data are generated by a pair of flow fields, one governing time evolution, the other a mapping between data and a low-dimensional latent space. Importantly, the dimension-reducing flow is trained to minimize distortions of the temporal dynamics, learning an identifiable low-dimensional representation that preserves temporal relations in the original data. Additionally, we constrain our latent spaces to have low-dimensional support in a soft, parameterized manner, taking inspiration from ideas on nested dropout. Across both neural and behavioral data, we show that this dual flow approach produces both more interpretable dynamics and higher-quality reconstructions than competing models, including in noise-dominated data sets where conventional approaches fail.
- Open Access
A synaptic locus of song learning
Summary: Learning by imitation is the foundation for verbal and musical expression, but its underlying neural basis remains obscure. A juvenile male zebra finch imitates the multisyllabic song of an adult tutor in a process that depends on a song-specialized cortico-basal ganglia circuit, affording a powerful system to identify the synaptic substrates of imitative motor learning. Plasticity at a particular set of cortico-basal ganglia synapses is hypothesized to drive rapid learning-related changes in song before these changes are subsequently consolidated in downstream circuits. Nevertheless, this hypothesis is untested and the synaptic locus where learning initially occurs is unknown. By combining a computational framework to quantify song learning with synapse-specific optogenetic and chemogenetic manipulations within and directly downstream of the cortico-basal ganglia circuit, we identified the specific cortico-basal ganglia synapses that drive the acquisition and expression of rapid vocal changes during juvenile song learning and characterized the hours-long timescale over which these changes consolidate. Furthermore, transiently augmenting postsynaptic activity in the basal ganglia briefly accelerates learning rates and persistently alters song, demonstrating a direct link between basal ganglia activity and rapid learning. These results localize the specific cortico-basal ganglia synapses that enable a juvenile songbird to learn to sing and reveal the circuit logic and behavioral timescales of this imitative learning paradigm.
Open AccessA synaptic locus of song learning
Summary: Learning by imitation is the foundation for verbal and musical expression, but its underlying neural basis remains obscure. A juvenile male zebra finch imitates the multisyllabic song of an adult tutor in a process that depends on a song-specialized cortico-basal ganglia circuit, affording a powerful system to identify the synaptic substrates of imitative motor learning. Plasticity at a particular set of cortico-basal ganglia synapses is hypothesized to drive rapid learning-related changes in song before these changes are subsequently consolidated in downstream circuits. Nevertheless, this hypothesis is untested and the synaptic locus where learning initially occurs is unknown. By combining a computational framework to quantify song learning with synapse-specific optogenetic and chemogenetic manipulations within and directly downstream of the cortico-basal ganglia circuit, we identified the specific cortico-basal ganglia synapses that drive the acquisition and expression of rapid vocal changes during juvenile song learning and characterized the hours-long timescale over which these changes consolidate. Furthermore, transiently augmenting postsynaptic activity in the basal ganglia briefly accelerates learning rates and persistently alters song, demonstrating a direct link between basal ganglia activity and rapid learning. These results localize the specific cortico-basal ganglia synapses that enable a juvenile songbird to learn to sing and reveal the circuit logic and behavioral timescales of this imitative learning paradigm.
- Open Access
Motor Cortical Output Integrates Distorted Proprioceptive Feedback
Proprioceptive feedback from muscles is essential for continuous monitoring and precise control of limb movement, yet how such peripheral feedback is integrated into ongoing descending motor cortical dynamics remains largely unclear. Here we used precisely timed optogenetic stimulation of forelimb muscles to manipulate proprioceptive signals during a mouse reach-to-consume task. Mice were able to successfully reach and consume despite muscle stimulation-evoked forelimb deviations. Across the cortex, the caudal forelimb area (CFA) was preferentially activated by muscle-specific proprioceptive inputs and contributed to stabilizing perturbed forelimb movement. CFA extratelencephalic (ET) neurons encoded movement kinematics as well as proprioceptive feedback. At neuron population level, proprioceptive perturbations rapidly deflected CFA ET neural trajectories away from the task-relevant manifold yet exerted only limited effects on evolving task dynamics. These results reveal that descending cortical circuit implements activity subspace separation to encode and integrate distorted proprioceptive feedback while preserving task-relevant output.
Open AccessMotor Cortical Output Integrates Distorted Proprioceptive Feedback
Proprioceptive feedback from muscles is essential for continuous monitoring and precise control of limb movement, yet how such peripheral feedback is integrated into ongoing descending motor cortical dynamics remains largely unclear. Here we used precisely timed optogenetic stimulation of forelimb muscles to manipulate proprioceptive signals during a mouse reach-to-consume task. Mice were able to successfully reach and consume despite muscle stimulation-evoked forelimb deviations. Across the cortex, the caudal forelimb area (CFA) was preferentially activated by muscle-specific proprioceptive inputs and contributed to stabilizing perturbed forelimb movement. CFA extratelencephalic (ET) neurons encoded movement kinematics as well as proprioceptive feedback. At neuron population level, proprioceptive perturbations rapidly deflected CFA ET neural trajectories away from the task-relevant manifold yet exerted only limited effects on evolving task dynamics. These results reveal that descending cortical circuit implements activity subspace separation to encode and integrate distorted proprioceptive feedback while preserving task-relevant output.
- Open Access
- 11 November 2025
- Published by Springer Nature in Nature Communications
- Vol. 16 (1) , 1-14
A software platform for real-time and adaptive neuroscience experiments
Current neuroscience research is often limited to testing predetermined hypotheses and post hoc analysis of already collected data. Adaptive experimental designs, in which modeling drives ongoing data collection and selects experimental manipulations, offer a promising alternative. However, such adaptive paradigms require tight integration between software and hardware under real-time constraints. We introduce improv , a software platform for flexible integration of modeling, data collection, analysis pipelines, and live experimental control. We demonstrate both in silico and in vivo how improv enables efficient experimental designs for discovery and validation across various model organisms and data types. We used improv to orchestrate real-time behavioral analyses, rapid functional typing of neural responses via calcium imaging, optimal visual stimulus selection, and model-driven optogenetic photostimulation of visually responsive neurons in the zebrafish brain. Together, these results demonstrate the power of improv to integrate modeling with data collection and experimental control to achieve next-generation adaptive experiments.
- 11 November 2025
- Published by Springer Nature in Nature Communications
- Vol. 16 (1) , 1-14
Open AccessA software platform for real-time and adaptive neuroscience experiments
Current neuroscience research is often limited to testing predetermined hypotheses and post hoc analysis of already collected data. Adaptive experimental designs, in which modeling drives ongoing data collection and selects experimental manipulations, offer a promising alternative. However, such adaptive paradigms require tight integration between software and hardware under real-time constraints. We introduce improv , a software platform for flexible integration of modeling, data collection, analysis pipelines, and live experimental control. We demonstrate both in silico and in vivo how improv enables efficient experimental designs for discovery and validation across various model organisms and data types. We used improv to orchestrate real-time behavioral analyses, rapid functional typing of neural responses via calcium imaging, optimal visual stimulus selection, and model-driven optogenetic photostimulation of visually responsive neurons in the zebrafish brain. Together, these results demonstrate the power of improv to integrate modeling with data collection and experimental control to achieve next-generation adaptive experiments.
- Open Access
Representational similarity and pattern classification of fifteen emotional states induced by movie clips and text scenarios
How do different emotional states relate to each other and how are they represented in the human brain? These are important questions in the field of affective neuroscience, with profound scientific and vast clinical implications. Different theories of emotion tend to emphasize the relative importance of distinct psychological constructs, for example categorical labels (e.g., fear, joy, or sadness) versus dimensional ratings (e.g., valence and arousal), for understanding human emotions. To investigate whether categorical or dimensional constructs correspond better to patterns of brain activity associated with human affective experiences, we experimentally induced 15 emotions (spanning positive, negative, and neutral valence) in 136 participants using 150 short movie clips and 150 one- or two-sentence text scenarios, while their blood oxygenation-level dependent activity was recorded in a magnetic resonance imaging scanner. Results from our representational similarity analyses suggest participants’ brain activity significantly correlated with their categorical labeling of, but not their dimensional rating of, the movie clips and text scenarios. Subsequently, we were also able to decode the categorical labels of these emotional stimuli using whole-brain multi-voxel pattern classification, with important voxels found in many cortical, limbic, subcortical, cerebellar, and brainstem regions. Finally, we found similar clusters of emotions through exploratory hierarchical clustering analyses of participants’ categorical labeling of and brain responses to these stimuli. Taken together, these findings greatly advance our understanding of how a large set of human emotions are related to each other both in terms of the participants’ self-report and their brain activity. Significance Statement: We successfully decoded fifteen emotional states induced by movie clips and text scenarios based on participants’ brain responses (blood oxygenation-level dependent signals) to these stimuli using multi-voxel pattern classification (a supervised machine learning approach). We also found that participants’ brain responses correlated with their self-reported emotional experience, i.e. which emotion they felt while they were presented with these stimuli, using representative similarity analysis (an unsupervised approach). Finally, we found that these emotions are organized in very similar ways both in terms of participants’ categorical labeling and their brain responses to these stimuli. Together, these data-driven and computational modeling-based findings greatly advance our understanding of how a large set of emotions are organized and represented in the human brain.
Open AccessRepresentational similarity and pattern classification of fifteen emotional states induced by movie clips and text scenarios
How do different emotional states relate to each other and how are they represented in the human brain? These are important questions in the field of affective neuroscience, with profound scientific and vast clinical implications. Different theories of emotion tend to emphasize the relative importance of distinct psychological constructs, for example categorical labels (e.g., fear, joy, or sadness) versus dimensional ratings (e.g., valence and arousal), for understanding human emotions. To investigate whether categorical or dimensional constructs correspond better to patterns of brain activity associated with human affective experiences, we experimentally induced 15 emotions (spanning positive, negative, and neutral valence) in 136 participants using 150 short movie clips and 150 one- or two-sentence text scenarios, while their blood oxygenation-level dependent activity was recorded in a magnetic resonance imaging scanner. Results from our representational similarity analyses suggest participants’ brain activity significantly correlated with their categorical labeling of, but not their dimensional rating of, the movie clips and text scenarios. Subsequently, we were also able to decode the categorical labels of these emotional stimuli using whole-brain multi-voxel pattern classification, with important voxels found in many cortical, limbic, subcortical, cerebellar, and brainstem regions. Finally, we found similar clusters of emotions through exploratory hierarchical clustering analyses of participants’ categorical labeling of and brain responses to these stimuli. Taken together, these findings greatly advance our understanding of how a large set of human emotions are related to each other both in terms of the participants’ self-report and their brain activity. Significance Statement: We successfully decoded fifteen emotional states induced by movie clips and text scenarios based on participants’ brain responses (blood oxygenation-level dependent signals) to these stimuli using multi-voxel pattern classification (a supervised machine learning approach). We also found that participants’ brain responses correlated with their self-reported emotional experience, i.e. which emotion they felt while they were presented with these stimuli, using representative similarity analysis (an unsupervised approach). Finally, we found that these emotions are organized in very similar ways both in terms of participants’ categorical labeling and their brain responses to these stimuli. Together, these data-driven and computational modeling-based findings greatly advance our understanding of how a large set of emotions are organized and represented in the human brain.
- Open Access
- 29 August 2025
- Published by Springer Nature in Nature
- Vol. 645 (8081) , 1
Author Correction: Dual neuromodulatory dynamics underlie birdsong learning
- 29 August 2025
- Published by Springer Nature in Nature
- Vol. 645 (8081) , 1
Open AccessAuthor Correction: Dual neuromodulatory dynamics underlie birdsong learning
- Open Access
Correctness is its own reward: bootstrapping error signals in self-guided reinforcement learning
Reinforcement learning (RL) offers a compelling account of how agents learn complex behaviors by trial and error, yet RL is predicated on the existence of a reward function provided by the agent’s environment. By contrast, many skills are learned without external guidance, posing a challenge to RL’s ability to account for self-directed learning. For instance, juvenile male zebra finches first memorize and then train themselves to reproduce the song of an adult male tutor through extensive practice. This process is believed to be guided by an internally computed assessment of performance quality, though the mechanism and development of this signal remain unknown. Here, we propose that, contrary to prevailing assumptions, tutor song memorization and performance assessment are subserved by the same neural circuit, one trained to predictively cancel tutor song. To test this hypothesis, we built models of a local forebrain circuit that uses contextual premotor signals to cancel tutor song auditory input via synaptic plasticity. After learning, excitatory projection neurons signaled mismatches between the tutor song and birds’ own performance, best matching experimental data when learning involved anti-Hebbian plasticity in recurrent interneurons. We also found that learning proceeds by both sharpening error sensitivity and minimizing responses to the tutor song. Finally, the error signals produced by this model can train a simple RL agent to replicate the spectrograms of adult bird songs. These results suggest that local learning via predictive cancellation suffices for bootstrapping error signals capable of guiding self-directed learning of natural behaviors. Significance: Many species, including humans, learn new skills by trial and error. But in the absence of external rewards and punishments, individuals need a means of assessing the quality of their performance. Here, we model the process by which juvenile male zebra finches memorize the song of an adult tutor, showing how local changes in brain circuits responsible for hearing can give rise during this process to an internal error signal that later allows for self-guided learning.
Open AccessCorrectness is its own reward: bootstrapping error signals in self-guided reinforcement learning
Reinforcement learning (RL) offers a compelling account of how agents learn complex behaviors by trial and error, yet RL is predicated on the existence of a reward function provided by the agent’s environment. By contrast, many skills are learned without external guidance, posing a challenge to RL’s ability to account for self-directed learning. For instance, juvenile male zebra finches first memorize and then train themselves to reproduce the song of an adult male tutor through extensive practice. This process is believed to be guided by an internally computed assessment of performance quality, though the mechanism and development of this signal remain unknown. Here, we propose that, contrary to prevailing assumptions, tutor song memorization and performance assessment are subserved by the same neural circuit, one trained to predictively cancel tutor song. To test this hypothesis, we built models of a local forebrain circuit that uses contextual premotor signals to cancel tutor song auditory input via synaptic plasticity. After learning, excitatory projection neurons signaled mismatches between the tutor song and birds’ own performance, best matching experimental data when learning involved anti-Hebbian plasticity in recurrent interneurons. We also found that learning proceeds by both sharpening error sensitivity and minimizing responses to the tutor song. Finally, the error signals produced by this model can train a simple RL agent to replicate the spectrograms of adult bird songs. These results suggest that local learning via predictive cancellation suffices for bootstrapping error signals capable of guiding self-directed learning of natural behaviors. Significance: Many species, including humans, learn new skills by trial and error. But in the absence of external rewards and punishments, individuals need a means of assessing the quality of their performance. Here, we model the process by which juvenile male zebra finches memorize the song of an adult tutor, showing how local changes in brain circuits responsible for hearing can give rise during this process to an internal error signal that later allows for self-guided learning.