Lin Zhong
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Research Subjects
Lin Zhong has a total of 38 co-authors. Here are the top co-authors by number of shared publications:
- Marius Pachitariu
HHMI Janelia Research Campus, Ashburn, USA
10 shared publications - Carsen Stringer
HHMI Janelia Research Campus, Ashburn, USA
9 shared publications - Atika Syeda
HHMI Janelia Research Campus, Ashburn, VA, USA
4 shared publications Alexa Gracias
HHMI Janelia Research Campus, Ashburn, USA
2 shared publicationsAmanda Minisi
HHMI Janelia Research Campus, Ashburn, USA
2 shared publicationsCrystall Lopez
HHMI Janelia Research Campus, Ashburn, USA
2 shared publications
Latest Publications
- 15 October 2026
- Published by American Association for the Advancement of Science (AAAS) in Science
- Vol. 394 (6821)
Dendritic computation for abstract rule-based flexible decision-making
Intelligent behavior involves flexible integration of sensory input and abstract rule knowledge. Dendritic computation has been shown to contribute to top-down modulation of sensory processing; however, whether and how it is involved in flexible cognitive processing remains to be elucidated. Combining a rule-based flexible decision-making task with in vivo two-photon imaging and optogenetics in mice, we found that within single layer 5 extratelencephalic neurons in the auditory cortex, apical dendrites preferentially encoded the inferred abstract rule and somata preferentially represented sensory stimuli, whereas choice signals were present in both compartments. Optogenetic dendritic inhibition impaired rule-dependent somatic choice coding, supporting a task-relevant integration across the dendritic and somatic compartments. These results indicate that dendritic computation in cortical neurons mediates flexible behavior.
- 15 October 2026
- Published by American Association for the Advancement of Science (AAAS) in Science
- Vol. 394 (6821)
Dendritic computation for abstract rule-based flexible decision-making
Intelligent behavior involves flexible integration of sensory input and abstract rule knowledge. Dendritic computation has been shown to contribute to top-down modulation of sensory processing; however, whether and how it is involved in flexible cognitive processing remains to be elucidated. Combining a rule-based flexible decision-making task with in vivo two-photon imaging and optogenetics in mice, we found that within single layer 5 extratelencephalic neurons in the auditory cortex, apical dendrites preferentially encoded the inferred abstract rule and somata preferentially represented sensory stimuli, whereas choice signals were present in both compartments. Optogenetic dendritic inhibition impaired rule-dependent somatic choice coding, supporting a task-relevant integration across the dendritic and somatic compartments. These results indicate that dendritic computation in cortical neurons mediates flexible behavior.
- Top Cited PapersOpen Access
- 20 May 2026
- Published by Springer Nature in Nature
- Vol. 655 (8124) , 1-7
A critical initialization for biological neural networks
Intrinsically generated, brainwide neural activity displays macroscopic coordination among large populations of neurons that persists beyond the biophysical timescales of individual neurons 1–3 . It is not well understood how these macroscopic behaviours arise from microscopic, short-lived interactions between pairs of neurons. Here we show that the eigenvalue spectrum and dynamical properties of large-scale neural recordings in mice are similar to those produced by linear dynamics governed by a random symmetric matrix that is critically normalized. An exception was population activity in hippocampal area CA1, which resembled an efficient, uncorrelated neural code that may be optimized for information storage capacity. High-dimensional, global activity modes emerged in critically normalized artificial networks and persisted under sparse, clustered or spatial connectivity. These dynamics were useful for solving time-dependent tasks such as a zero-shot working memory task.
- 20 May 2026
- Published by Springer Nature in Nature
- Vol. 655 (8124) , 1-7
Top Cited PapersOpen AccessA critical initialization for biological neural networks
Intrinsically generated, brainwide neural activity displays macroscopic coordination among large populations of neurons that persists beyond the biophysical timescales of individual neurons 1–3 . It is not well understood how these macroscopic behaviours arise from microscopic, short-lived interactions between pairs of neurons. Here we show that the eigenvalue spectrum and dynamical properties of large-scale neural recordings in mice are similar to those produced by linear dynamics governed by a random symmetric matrix that is critically normalized. An exception was population activity in hippocampal area CA1, which resembled an efficient, uncorrelated neural code that may be optimized for information storage capacity. High-dimensional, global activity modes emerged in critically normalized artificial networks and persisted under sparse, clustered or spatial connectivity. These dynamics were useful for solving time-dependent tasks such as a zero-shot working memory task.
- Open Access
Linking neural representations to behavior using generalization
Sensory-guided decisions are the result of sensorimotor transformations across many brain areas. Recent studies have localized the motor- and decision-related components of these transformations using brain-wide neural recordings. It has been more difficult to localize sensory computations in the same way. Here we developed a new approach for linking sensory computations to behavior by training mice to discriminate between two stimuli and testing their responses with new stimuli. In separate animals, we calculated the similarity of neural representations between train and test stimuli, using recordings of up to 73,000 simultaneously-recorded neurons from 9 primary and higher-order visual areas (HVAs) across layers 2 and 3. We found that neural discrimination on test but not train images correlated with behavioral discrimination, and this relation required prior visual experience as it was not present in dark-reared mice. The link between neural and behavioral performance was highest in the medial HVAs, suggesting this region is a critical component of sensory transformations and generalization.
Open AccessLinking neural representations to behavior using generalization
Sensory-guided decisions are the result of sensorimotor transformations across many brain areas. Recent studies have localized the motor- and decision-related components of these transformations using brain-wide neural recordings. It has been more difficult to localize sensory computations in the same way. Here we developed a new approach for linking sensory computations to behavior by training mice to discriminate between two stimuli and testing their responses with new stimuli. In separate animals, we calculated the similarity of neural representations between train and test stimuli, using recordings of up to 73,000 simultaneously-recorded neurons from 9 primary and higher-order visual areas (HVAs) across layers 2 and 3. We found that neural discrimination on test but not train images correlated with behavioral discrimination, and this relation required prior visual experience as it was not present in dark-reared mice. The link between neural and behavioral performance was highest in the medial HVAs, suggesting this region is a critical component of sensory transformations and generalization.
- Open Access
Orofacial behaviors, not eye movements, govern neural activity in mouse visual cortex
Neural activity in mouse primary visual cortex (V1) correlates strongly with orofacial movements. Such strong modulation has not been found in the primate visual cortex during eye fixation [1], which led to the suggestion that the modulation may primarily depend on eye movements in both species [2]. Here we examined the influence of eye movements on neural activity in mouse visual cortex both in complete darkness and in the presence of different types of visual input. In all cases, we found that eye movements explain a smaller fraction of neural activity variance compared to orofacial behaviors. Additionally, we found that eye movements were correlated to orofacial movements, such as whisking and sniffing, and thus may be indirectly correlated to neural activity. These results further emphasize the impact of movement signals on mouse visual cortex during free viewing behavior.
Open AccessOrofacial behaviors, not eye movements, govern neural activity in mouse visual cortex
Neural activity in mouse primary visual cortex (V1) correlates strongly with orofacial movements. Such strong modulation has not been found in the primate visual cortex during eye fixation [1], which led to the suggestion that the modulation may primarily depend on eye movements in both species [2]. Here we examined the influence of eye movements on neural activity in mouse visual cortex both in complete darkness and in the presence of different types of visual input. In all cases, we found that eye movements explain a smaller fraction of neural activity variance compared to orofacial behaviors. Additionally, we found that eye movements were correlated to orofacial movements, such as whisking and sniffing, and thus may be indirectly correlated to neural activity. These results further emphasize the impact of movement signals on mouse visual cortex during free viewing behavior.
- Open Access
- 18 June 2025
- Published by Springer Nature in Nature
- Vol. 644 (8077) , 741-748
Unsupervised pretraining in biological neural networks
Representation learning in neural networks may be implemented with supervised or unsupervised algorithms, distinguished by the availability of instruction. In the sensory cortex, perceptual learning drives neural plasticity1–13, but it is not known whether this is due to supervised or unsupervised learning. Here we recorded populations of up to 90,000 neurons simultaneously from the primary visual cortex (V1) and higher visual areas (HVAs) while mice learned multiple tasks, as well as during unrewarded exposure to the same stimuli. Similar to previous studies, we found that neural changes in task mice were correlated with their behavioural learning. However, the neural changes were mostly replicated in mice with unrewarded exposure, suggesting that the changes were in fact due to unsupervised learning. The neural plasticity was highest in the medial HVAs and obeyed visual, rather than spatial, learning rules. In task mice only, we found a ramping reward-prediction signal in anterior HVAs, potentially involved in supervised learning. Our neural results predict that unsupervised learning may accelerate subsequent task learning, a prediction that we validated with behavioural experiments.
- 18 June 2025
- Published by Springer Nature in Nature
- Vol. 644 (8077) , 741-748
Open AccessUnsupervised pretraining in biological neural networks
Representation learning in neural networks may be implemented with supervised or unsupervised algorithms, distinguished by the availability of instruction. In the sensory cortex, perceptual learning drives neural plasticity1–13, but it is not known whether this is due to supervised or unsupervised learning. Here we recorded populations of up to 90,000 neurons simultaneously from the primary visual cortex (V1) and higher visual areas (HVAs) while mice learned multiple tasks, as well as during unrewarded exposure to the same stimuli. Similar to previous studies, we found that neural changes in task mice were correlated with their behavioural learning. However, the neural changes were mostly replicated in mice with unrewarded exposure, suggesting that the changes were in fact due to unsupervised learning. The neural plasticity was highest in the medial HVAs and obeyed visual, rather than spatial, learning rules. In task mice only, we found a ramping reward-prediction signal in anterior HVAs, potentially involved in supervised learning. Our neural results predict that unsupervised learning may accelerate subsequent task learning, a prediction that we validated with behavioural experiments.
- Open Access
A critical initialization for biological neural networks
Artificial neural networks learn faster if they are initialized well. Good initializations can generate high-dimensional macroscopic dynamics with long timescales. It is not known if biological neural networks have similar properties. Here we show that the eigenvalue spectrum and dynamical properties of large-scale neural recordings in mice (two-photon and electrophysiology) are similar to those produced by linear dynamics governed by a random symmetric matrix that is critically normalized. An exception was hippocampal area CA1: population activity in this area resembled an efficient, uncorrelated neural code, which may be optimized for information storage capacity. Global emergent activity modes persisted in simulations with sparse, clustered or spatial connectivity. We hypothesize that the spontaneous neural activity reflects a critical initialization of whole-brain neural circuits that is optimized for learning time-dependent tasks.
Open AccessA critical initialization for biological neural networks
Artificial neural networks learn faster if they are initialized well. Good initializations can generate high-dimensional macroscopic dynamics with long timescales. It is not known if biological neural networks have similar properties. Here we show that the eigenvalue spectrum and dynamical properties of large-scale neural recordings in mice (two-photon and electrophysiology) are similar to those produced by linear dynamics governed by a random symmetric matrix that is critically normalized. An exception was hippocampal area CA1: population activity in this area resembled an efficient, uncorrelated neural code, which may be optimized for information storage capacity. Global emergent activity modes persisted in simulations with sparse, clustered or spatial connectivity. We hypothesize that the spontaneous neural activity reflects a critical initialization of whole-brain neural circuits that is optimized for learning time-dependent tasks.
- Top Cited PapersOpen Access
- 16 October 2024
- Published by Springer Nature in Nature Neuroscience
- Vol. 28 (1) , 201-212
Rastermap: a discovery method for neural population recordings
Neurophysiology has long progressed through exploratory experiments and chance discoveries. Anecdotes abound of researchers listening to spikes in real time and noticing patterns of activity related to ongoing stimuli or behaviors. With the advent of large-scale recordings, such close observation of data has become difficult. To find patterns in large-scale neural data, we developed ‘Rastermap’, a visualization method that displays neurons as a raster plot after sorting them along a one-dimensional axis based on their activity patterns. We benchmarked Rastermap on realistic simulations and then used it to explore recordings of tens of thousands of neurons from mouse cortex during spontaneous, stimulus-evoked and task-evoked epochs. We also applied Rastermap to whole-brain zebrafish recordings; to wide-field imaging data; to electrophysiological recordings in rat hippocampus, monkey frontal cortex and various cortical and subcortical regions in mice; and to artificial neural networks. Finally, we illustrate high-dimensional scenarios where Rastermap and similar algorithms cannot be used effectively.
- 16 October 2024
- Published by Springer Nature in Nature Neuroscience
- Vol. 28 (1) , 201-212
Top Cited PapersOpen AccessRastermap: a discovery method for neural population recordings
Neurophysiology has long progressed through exploratory experiments and chance discoveries. Anecdotes abound of researchers listening to spikes in real time and noticing patterns of activity related to ongoing stimuli or behaviors. With the advent of large-scale recordings, such close observation of data has become difficult. To find patterns in large-scale neural data, we developed ‘Rastermap’, a visualization method that displays neurons as a raster plot after sorting them along a one-dimensional axis based on their activity patterns. We benchmarked Rastermap on realistic simulations and then used it to explore recordings of tens of thousands of neurons from mouse cortex during spontaneous, stimulus-evoked and task-evoked epochs. We also applied Rastermap to whole-brain zebrafish recordings; to wide-field imaging data; to electrophysiological recordings in rat hippocampus, monkey frontal cortex and various cortical and subcortical regions in mice; and to artificial neural networks. Finally, we illustrate high-dimensional scenarios where Rastermap and similar algorithms cannot be used effectively.
- Open Access
Equivariant neuronal populations enable simultaneous tuning and invariance
As we move through the world, we see the same visual scene from different perspectives. But how does the brain encode scene identity invariant to perspective, while remaining sensitive to these transformations? We propose a solution through equivariance, where perspective transformations induce structured changes in neuronal population responses. This framework implies a decomposition of population responses into orthogonal subspaces that are tuned and invariant. Testing our framework with large-scale neuronal recordings across four mouse visual cortical areas, we find that the equivariant structure is more pronounced in some higher-order areas (LM, AL) than in other areas (V1, RL). This equivariant structure accounts for the observed simultaneous increase in both population tuning and invariance. In comparison, early layers of an artificial neural network trained on image classification show similar structure, but later layers increase invariance at the cost of tuning. These results suggest equivariance is a principle to achieve flexible computations with neuronal populations.
Open AccessEquivariant neuronal populations enable simultaneous tuning and invariance
As we move through the world, we see the same visual scene from different perspectives. But how does the brain encode scene identity invariant to perspective, while remaining sensitive to these transformations? We propose a solution through equivariance, where perspective transformations induce structured changes in neuronal population responses. This framework implies a decomposition of population responses into orthogonal subspaces that are tuned and invariant. Testing our framework with large-scale neuronal recordings across four mouse visual cortical areas, we find that the equivariant structure is more pronounced in some higher-order areas (LM, AL) than in other areas (V1, RL). This equivariant structure accounts for the observed simultaneous increase in both population tuning and invariance. In comparison, early layers of an artificial neural network trained on image classification show similar structure, but later layers increase invariance at the cost of tuning. These results suggest equivariance is a principle to achieve flexible computations with neuronal populations.
- Open Access
Distinct streams for supervised and unsupervised learning in the visual cortex
Representation learning in neural networks may be implemented with supervised or unsupervised algorithms, distinguished by the availability of feedback. In sensory cortex, perceptual learning drives neural plasticity, but it is not known if this is due to supervised or unsupervised learning. Here we recorded populations of up to 90,000 neurons simultaneously from the primary visual cortex (V1) and higher visual areas (HVA), while mice learned multiple tasks as well as during unrewarded exposure to the same stimuli. Similar to previous studies, we found that neural changes in task mice were correlated with their behavioral learning. However, the neural changes were mostly replicated in mice with unrewarded exposure, suggesting that the changes were in fact due to unsupervised learning. The neural plasticity was concentrated in the medial HVAs and obeyed visual, rather than spatial, learning rules. In task mice only, we found a ramping reward prediction signal in anterior HVAs, potentially involved in supervised learning. Our neural results predict that unsupervised learning may accelerate subsequent task learning, a prediction which we validated with behavioral experiments.
Open AccessDistinct streams for supervised and unsupervised learning in the visual cortex
Representation learning in neural networks may be implemented with supervised or unsupervised algorithms, distinguished by the availability of feedback. In sensory cortex, perceptual learning drives neural plasticity, but it is not known if this is due to supervised or unsupervised learning. Here we recorded populations of up to 90,000 neurons simultaneously from the primary visual cortex (V1) and higher visual areas (HVA), while mice learned multiple tasks as well as during unrewarded exposure to the same stimuli. Similar to previous studies, we found that neural changes in task mice were correlated with their behavioral learning. However, the neural changes were mostly replicated in mice with unrewarded exposure, suggesting that the changes were in fact due to unsupervised learning. The neural plasticity was concentrated in the medial HVAs and obeyed visual, rather than spatial, learning rules. In task mice only, we found a ramping reward prediction signal in anterior HVAs, potentially involved in supervised learning. Our neural results predict that unsupervised learning may accelerate subsequent task learning, a prediction which we validated with behavioral experiments.
- Top Cited PapersOpen Access
- 20 November 2023
- Published by Springer Nature in Nature Neuroscience
- Vol. 27 (1) , 187-195
Facemap: a framework for modeling neural activity based on orofacial tracking
Recent studies in mice have shown that orofacial behaviors drive a large fraction of neural activity across the brain. To understand the nature and function of these signals, we need better computational models to characterize the behaviors and relate them to neural activity. Here we developed Facemap, a framework consisting of a keypoint tracker and a deep neural network encoder for predicting neural activity. Our algorithm for tracking mouse orofacial behaviors was more accurate than existing pose estimation tools, while the processing speed was several times faster, making it a powerful tool for real-time experimental interventions. The Facemap tracker was easy to adapt to data from new labs, requiring as few as 10 annotated frames for near-optimal performance. We used the keypoints as inputs to a deep neural network which predicts the activity of ~50,000 simultaneously-recorded neurons and, in visual cortex, we doubled the amount of explained variance compared to previous methods. Using this model, we found that the neuronal activity clusters that were well predicted from behavior were more spatially spread out across cortex. We also found that the deep behavioral features from the model had stereotypical, sequential dynamics that were not reversible in time. In summary, Facemap provides a stepping stone toward understanding the function of the brain-wide neural signals and their relation to behavior.
- 20 November 2023
- Published by Springer Nature in Nature Neuroscience
- Vol. 27 (1) , 187-195
Top Cited PapersOpen AccessFacemap: a framework for modeling neural activity based on orofacial tracking
Recent studies in mice have shown that orofacial behaviors drive a large fraction of neural activity across the brain. To understand the nature and function of these signals, we need better computational models to characterize the behaviors and relate them to neural activity. Here we developed Facemap, a framework consisting of a keypoint tracker and a deep neural network encoder for predicting neural activity. Our algorithm for tracking mouse orofacial behaviors was more accurate than existing pose estimation tools, while the processing speed was several times faster, making it a powerful tool for real-time experimental interventions. The Facemap tracker was easy to adapt to data from new labs, requiring as few as 10 annotated frames for near-optimal performance. We used the keypoints as inputs to a deep neural network which predicts the activity of ~50,000 simultaneously-recorded neurons and, in visual cortex, we doubled the amount of explained variance compared to previous methods. Using this model, we found that the neuronal activity clusters that were well predicted from behavior were more spatially spread out across cortex. We also found that the deep behavioral features from the model had stereotypical, sequential dynamics that were not reversible in time. In summary, Facemap provides a stepping stone toward understanding the function of the brain-wide neural signals and their relation to behavior.