Ling-Qi Zhang
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
Research Interests
Ling-Qi Zhang has a total of 36 co-authors. Here are the top co-authors by number of shared publications:
- Alan A. Stocker
Department of Psychology, University of Pennsylvania
5 shared publications - David H. Brainard
Department of Psychology, University of Pennsylvania, Philadelphia, PA, USA
3 shared publications - Nicolas P. Cottaris
University of Pennsylvania
3 shared publications Cheng Qiu
Department of Psychology, University of Pennsylvania, Philadelphia, USA
2 shared publications- Geoffrey Karl Aguirre
University of Pennsylvania
2 shared publications - Jiang Mao
Department of Psychology, University of Pennsylvania
2 shared publications
Latest Publications
- Open Access
Inductive bias anchors a scaling law for biological learning
The speed of animal learning varies substantially across individuals and task domains, and does not always track intuitive notions of task complexity. A nominally simple binary decision task can take thousands of rewards to reach proficiency, whereas seemingly complex naturalistic tasks can be learned within dozens. A long-standing explanation is that learning speed reflects the alignment between an animal’s inductive biases and the task at hand. However, without a precise notion of alignment, it is difficult to translate this idea into quantitative predictions. Here, modeling an animal as a Bayesian learner that improves through iterative reward feedback, we use statistical learning theory to derive a scaling law for biological learning. The law makes two predictions. First, inductive bias multiplicatively sets the speed of learning through a single quantity, the effective complexity . This quantity measures how well the task aligns with the animal’s prior, and corresponds to the number of bits required to represent the task solution under that prior. We demonstrate this multiplicative scaling in a human motor learning dataset, where behavioral measurements before learning can precisely predict subsequent learning speed. Second, task performance improves as a power law in experience. We validate this power law in a large-scale mouse learning dataset and show that a two-parameter rescaling collapses individual learning curves onto a universal function predicted by the theory. These results suggest that behavioral measurements before learning can be used to predict task-dependent learning speeds, and they explain why learning speed reflects alignment with prior knowledge rather than nominal task complexity.
Open AccessInductive bias anchors a scaling law for biological learning
The speed of animal learning varies substantially across individuals and task domains, and does not always track intuitive notions of task complexity. A nominally simple binary decision task can take thousands of rewards to reach proficiency, whereas seemingly complex naturalistic tasks can be learned within dozens. A long-standing explanation is that learning speed reflects the alignment between an animal’s inductive biases and the task at hand. However, without a precise notion of alignment, it is difficult to translate this idea into quantitative predictions. Here, modeling an animal as a Bayesian learner that improves through iterative reward feedback, we use statistical learning theory to derive a scaling law for biological learning. The law makes two predictions. First, inductive bias multiplicatively sets the speed of learning through a single quantity, the effective complexity . This quantity measures how well the task aligns with the animal’s prior, and corresponds to the number of bits required to represent the task solution under that prior. We demonstrate this multiplicative scaling in a human motor learning dataset, where behavioral measurements before learning can precisely predict subsequent learning speed. Second, task performance improves as a power law in experience. We validate this power law in a large-scale mouse learning dataset and show that a two-parameter rescaling collapses individual learning curves onto a universal function predicted by the theory. These results suggest that behavioral measurements before learning can be used to predict task-dependent learning speeds, and they explain why learning speed reflects alignment with prior knowledge rather than nominal task complexity.
- Open Access
Mus musculus mice self-organize existing behaviors into structured bouts as they become expert hunters
SUMMARY: In the wild, a mouse must flexibly perform multiple computations at once, rapidly navigating, sampling its environment, forming and using memories, and balancing internal needs. Capturing the self-paced, low-repetition, complex nature of decision making in the wild while preserving the experimental control needed to interpret its process remains a challenge. Here we show that lab mice, without food restriction, learn to forage for hidden resources cued by ambiguous sounds in a dynamic environment by spatiotemporally reorganizing their existing behavioral repertoire. Using a novel closed-loop assay, we required mice to sit still to receive informative but hard-to-localize sound cues emanating from one correct location they must find, out of 157 possible locations in a large arena. With experience, mice become efficient hunters. Efficient hunting cannot be explained by an increase in a particular behavioral module, but instead, mice reorganize existing behaviors into clustered bouts of high-quality sampling and site-checking. Looking forward, this work establishes an approach for studying the neural, molecular, and evolutionary basis of naturalistic decision-making in mice.
Open AccessMus musculus mice self-organize existing behaviors into structured bouts as they become expert hunters
SUMMARY: In the wild, a mouse must flexibly perform multiple computations at once, rapidly navigating, sampling its environment, forming and using memories, and balancing internal needs. Capturing the self-paced, low-repetition, complex nature of decision making in the wild while preserving the experimental control needed to interpret its process remains a challenge. Here we show that lab mice, without food restriction, learn to forage for hidden resources cued by ambiguous sounds in a dynamic environment by spatiotemporally reorganizing their existing behavioral repertoire. Using a novel closed-loop assay, we required mice to sit still to receive informative but hard-to-localize sound cues emanating from one correct location they must find, out of 157 possible locations in a large arena. With experience, mice become efficient hunters. Efficient hunting cannot be explained by an increase in a particular behavioral module, but instead, mice reorganize existing behaviors into clustered bouts of high-quality sampling and site-checking. Looking forward, this work establishes an approach for studying the neural, molecular, and evolutionary basis of naturalistic decision-making in mice.
- Open Access
Transformation-tolerant object recognition in tree shrews despite lacking a fovea
Object recognition depends on the ability to extract stable representations across changes in how they are viewed, yet it remains unclear how this capacity depends on visual acuity and cortical hierarchy. We combined behavioral testing and computational modeling to determine whether tree shrews, close relatives of primates with lower spatial acuity, can perform transformation-tolerant object recognition. Front-end modeling incorporating species-specific optics and photoreceptor sampling showed that, when scaled for acuity, tree shrew retinal filtering preserves the similarity structure of natural image categories relevant for object recognition. Behaviorally, tree shrews reliably discriminated complex objects across variations in position, scale, and viewpoint, including when embedded within natural scenes, and generalized to novel exemplars. Their recognition behavior was best explained by visual features emphasizing differences in global shape and size between objects and by representations from intermediate and deep layers of hierarchical neural network models. These results demonstrate that visual processing supporting object-level generalization can arise within visual systems lacking high-acuity front-end optics and establish the tree shrew as a key model for understanding the computational and evolutionary origins of high-level vision.
Open AccessTransformation-tolerant object recognition in tree shrews despite lacking a fovea
Object recognition depends on the ability to extract stable representations across changes in how they are viewed, yet it remains unclear how this capacity depends on visual acuity and cortical hierarchy. We combined behavioral testing and computational modeling to determine whether tree shrews, close relatives of primates with lower spatial acuity, can perform transformation-tolerant object recognition. Front-end modeling incorporating species-specific optics and photoreceptor sampling showed that, when scaled for acuity, tree shrew retinal filtering preserves the similarity structure of natural image categories relevant for object recognition. Behaviorally, tree shrews reliably discriminated complex objects across variations in position, scale, and viewpoint, including when embedded within natural scenes, and generalized to novel exemplars. Their recognition behavior was best explained by visual features emphasizing differences in global shape and size between objects and by representations from intermediate and deep layers of hierarchical neural network models. These results demonstrate that visual processing supporting object-level generalization can arise within visual systems lacking high-acuity front-end optics and establish the tree shrew as a key model for understanding the computational and evolutionary origins of high-level vision.
- Open Access
- 23 April 2025
- Published by National Academy of Sciences in Proceedings of the National Academy of Sciences
- Vol. 122 (17)
The tilt illusion arises from an efficient reallocation of neural coding resources at the contextual boundary
The tilt illusion—a bias in the perceived orientation of a center stimulus induced by an oriented surround—illustrates how context shapes visual perception. Although extensively studied for decades, we still lack a comprehensive account of the illusion that connects its behavioral and neural characteristics. Here, we demonstrate that the tilt illusion originates from dynamic changes in neural coding precision induced by the surround context. We simultaneously obtained psychophysical and functional MRI responses from human subjects while they viewed gratings in the absence and presence of an oriented surround and independently extracted sensory encoding precision from their behavioral and neural data. Both measures show that in the absence of an oriented surround, encoding reflects the natural scene statistics of orientation. However, with an oriented surround, encoding precision is significantly increased for stimuli similar to the surround orientation. This local change in encoding is sufficient to predict the behavioral characteristics of the tilt illusion using a Bayesian observer model. The effect of surround modulation increases along the ventral stream and is localized to the portion of the visual cortex with receptive fields at the center-surround boundary. The pattern of change in coding accuracy reflects the surround-conditioned orientation statistics in natural scenes, but cannot be explained by local stimulus configuration. Our results suggest that the tilt illusion naturally emerges from an adaptive coding strategy that efficiently reallocates neural coding resources based on the current stimulus context.
- 23 April 2025
- Published by National Academy of Sciences in Proceedings of the National Academy of Sciences
- Vol. 122 (17)
Open AccessThe tilt illusion arises from an efficient reallocation of neural coding resources at the contextual boundary
The tilt illusion—a bias in the perceived orientation of a center stimulus induced by an oriented surround—illustrates how context shapes visual perception. Although extensively studied for decades, we still lack a comprehensive account of the illusion that connects its behavioral and neural characteristics. Here, we demonstrate that the tilt illusion originates from dynamic changes in neural coding precision induced by the surround context. We simultaneously obtained psychophysical and functional MRI responses from human subjects while they viewed gratings in the absence and presence of an oriented surround and independently extracted sensory encoding precision from their behavioral and neural data. Both measures show that in the absence of an oriented surround, encoding reflects the natural scene statistics of orientation. However, with an oriented surround, encoding precision is significantly increased for stimuli similar to the surround orientation. This local change in encoding is sufficient to predict the behavioral characteristics of the tilt illusion using a Bayesian observer model. The effect of surround modulation increases along the ventral stream and is localized to the portion of the visual cortex with receptive fields at the center-surround boundary. The pattern of change in coding accuracy reflects the surround-conditioned orientation statistics in natural scenes, but cannot be explained by local stimulus configuration. Our results suggest that the tilt illusion naturally emerges from an adaptive coding strategy that efficiently reallocates neural coding resources based on the current stimulus context.
- Open Access
The tilt illusion arises from an efficient reallocation of neural coding resources at the contextual boundary
The tilt illusion — a bias in the perceived orientation of a center stimulus induced by an oriented surround — illustrates how context shapes visual perception. While the tilt illusion has been the subject of quantitative study for over 85 years, we still lack a comprehensive account of the phenomenon that connects its neural and behavioral characteristics. Here, we demonstrate that the tilt illusion originates from a dynamic change in neural coding precision induced by the surround context. We simultaneously obtained psychophysical and fMRI responses from human subjects while they viewed gratings in the absence and presence of an oriented surround, and extracted sensory encoding precision from their behavioral and neural data. Both measures show that in the absence of a surround, encoding reflects the natural scene statistics of orientation. However, in the presence of an oriented surround, encoding precision is significantly increased for stimuli similar to the surround orientation. This local change in encoding is sufficient to accurately predict the behavioral characteristics of the tilt illusion using a Bayesian observer model. The effect of surround modulation increases along the ventral stream, and is localized to the portion of the visual cortex with receptive fields at the center-surround boundary. The pattern of change in coding accuracy reflects the surround-conditioned orientation statistics in natural scenes, but cannot be explained by local stimulus configuration. Our results suggest that the tilt illusion naturally emerges from a dynamic coding strategy that efficiently reallocates neural coding resources based on the current stimulus context.
Open AccessThe tilt illusion arises from an efficient reallocation of neural coding resources at the contextual boundary
The tilt illusion — a bias in the perceived orientation of a center stimulus induced by an oriented surround — illustrates how context shapes visual perception. While the tilt illusion has been the subject of quantitative study for over 85 years, we still lack a comprehensive account of the phenomenon that connects its neural and behavioral characteristics. Here, we demonstrate that the tilt illusion originates from a dynamic change in neural coding precision induced by the surround context. We simultaneously obtained psychophysical and fMRI responses from human subjects while they viewed gratings in the absence and presence of an oriented surround, and extracted sensory encoding precision from their behavioral and neural data. Both measures show that in the absence of a surround, encoding reflects the natural scene statistics of orientation. However, in the presence of an oriented surround, encoding precision is significantly increased for stimuli similar to the surround orientation. This local change in encoding is sufficient to accurately predict the behavioral characteristics of the tilt illusion using a Bayesian observer model. The effect of surround modulation increases along the ventral stream, and is localized to the portion of the visual cortex with receptive fields at the center-surround boundary. The pattern of change in coding accuracy reflects the surround-conditioned orientation statistics in natural scenes, but cannot be explained by local stimulus configuration. Our results suggest that the tilt illusion naturally emerges from a dynamic coding strategy that efficiently reallocates neural coding resources based on the current stimulus context.
- 16 January 2023
- Published by Society for Imaging Science & Technology in Electronic Imaging
- Vol. 35 (8) , 310-1
Towards an image-computable visual text quality metric using deep neural networks
Image quality metrics have become invaluable tools for image processing and display system development. These metrics are typically developed for and tested on images and videos of natural content. Text, on the other hand, has unique features and supports a distinct visual function: reading. It is therefore not clear if these image quality metrics are efficient or optimal as measures of text quality. Here, we developed a domain-specific image quality metric for text and compared its performance against quality metrics developed for natural images. To develop our metric, we first trained a deep neural network to perform text classification on a data set of distorted letter images. We then compute the responses of internal layers of the network to uncorrupted and corrupted images of text, respectively. We used the cosine dissimilarity between the responses as a measure of text quality. Preliminary results indicate that both our model and more established quality metrics (e.g., SSIM) are able to predict general trends in participants text quality ratings. In some cases, our model is able to outperform SSIM. We further developed our model to predict response data in a two-alternative forced choice experiment, on which only our model achieved very high accuracy.
- 16 January 2023
- Published by Society for Imaging Science & Technology in Electronic Imaging
- Vol. 35 (8) , 310-1
Towards an image-computable visual text quality metric using deep neural networks
Image quality metrics have become invaluable tools for image processing and display system development. These metrics are typically developed for and tested on images and videos of natural content. Text, on the other hand, has unique features and supports a distinct visual function: reading. It is therefore not clear if these image quality metrics are efficient or optimal as measures of text quality. Here, we developed a domain-specific image quality metric for text and compared its performance against quality metrics developed for natural images. To develop our metric, we first trained a deep neural network to perform text classification on a data set of distorted letter images. We then compute the responses of internal layers of the network to uncorrupted and corrupted images of text, respectively. We used the cosine dissimilarity between the responses as a measure of text quality. Preliminary results indicate that both our model and more established quality metrics (e.g., SSIM) are able to predict general trends in participants text quality ratings. In some cases, our model is able to outperform SSIM. We further developed our model to predict response data in a two-alternative forced choice experiment, on which only our model achieved very high accuracy.
- Open Access
- 29 December 2022
- Published by Springer Nature in Nature Communications
- Vol. 13 (1) , 1-12
Efficient neural codes naturally emerge through gradient descent learning
Human sensory systems are more sensitive to common features in the environment than uncommon features. For example, small deviations from the more frequently encountered horizontal orientations can be more easily detected than small deviations from the less frequent diagonal ones. Here we find that artificial neural networks trained to recognize objects also have patterns of sensitivity that match the statistics of features in images. To interpret these findings, we show mathematically that learning with gradient descent in neural networks preferentially creates representations that are more sensitive to common features, a hallmark of efficient coding. This effect occurs in systems with otherwise unconstrained coding resources, and additionally when learning towards both supervised and unsupervised objectives. This result demonstrates that efficient codes can naturally emerge from gradient-like learning.
- 29 December 2022
- Published by Springer Nature in Nature Communications
- Vol. 13 (1) , 1-12
Open AccessEfficient neural codes naturally emerge through gradient descent learning
Human sensory systems are more sensitive to common features in the environment than uncommon features. For example, small deviations from the more frequently encountered horizontal orientations can be more easily detected than small deviations from the less frequent diagonal ones. Here we find that artificial neural networks trained to recognize objects also have patterns of sensitivity that match the statistics of features in images. To interpret these findings, we show mathematically that learning with gradient descent in neural networks preferentially creates representations that are more sensitive to common features, a hallmark of efficient coding. This effect occurs in systems with otherwise unconstrained coding resources, and additionally when learning towards both supervised and unsupervised objectives. This result demonstrates that efficient codes can naturally emerge from gradient-like learning.
- Open Access
- 17 January 2022
- Published by eLife Sciences Publications, Ltd in eLife
- Vol. 11
An image reconstruction framework for characterizing initial visual encoding
We developed an image-computable observer model of the initial visual encoding that operates on natural image input, based on the framework of Bayesian image reconstruction from the excitations of the retinal cone mosaic. Our model extends previous work on ideal observer analysis and evaluation of performance beyond psychophysical discrimination, takes into account the statistical regularities of the visual environment, and provides a unifying framework for answering a wide range of questions regarding the visual front end. Using the error in the reconstructions as a metric, we analyzed variations of the number of different photoreceptor types on human retina as an optimal design problem. In addition, the reconstructions allow both visualization and quantification of information loss due to physiological optics and cone mosaic sampling, and how these vary with eccentricity. Furthermore, in simulations of color deficiencies and interferometric experiments, we found that the reconstructed images provide a reasonable proxy for modeling subjects’ percepts. Lastly, we used the reconstruction-based observer for the analysis of psychophysical threshold, and found notable interactions between spatial frequency and chromatic direction in the resulting spatial contrast sensitivity function. Our method is widely applicable to experiments and applications in which the initial visual encoding plays an important role.
- 17 January 2022
- Published by eLife Sciences Publications, Ltd in eLife
- Vol. 11
Open AccessAn image reconstruction framework for characterizing initial visual encoding
We developed an image-computable observer model of the initial visual encoding that operates on natural image input, based on the framework of Bayesian image reconstruction from the excitations of the retinal cone mosaic. Our model extends previous work on ideal observer analysis and evaluation of performance beyond psychophysical discrimination, takes into account the statistical regularities of the visual environment, and provides a unifying framework for answering a wide range of questions regarding the visual front end. Using the error in the reconstructions as a metric, we analyzed variations of the number of different photoreceptor types on human retina as an optimal design problem. In addition, the reconstructions allow both visualization and quantification of information loss due to physiological optics and cone mosaic sampling, and how these vary with eccentricity. Furthermore, in simulations of color deficiencies and interferometric experiments, we found that the reconstructed images provide a reasonable proxy for modeling subjects’ percepts. Lastly, we used the reconstruction-based observer for the analysis of psychophysical threshold, and found notable interactions between spatial frequency and chromatic direction in the resulting spatial contrast sensitivity function. Our method is widely applicable to experiments and applications in which the initial visual encoding plays an important role.
- Open Access
Prior expectations in visual speed perception predict encoding characteristics of neurons in area MT
Bayesian inference provides an elegant theoretical framework for understanding the characteristic biases and discrimination thresholds in visual speed perception. However, the framework is difficult to validate due to its flexibility and the fact that suitable constraints on the structure of the sensory uncertainty have been missing. Here, we demonstrate that a Bayesian observer model constrained by efficient coding not only well explains human visual speed perception but also provides an accurate quantitative account of the tuning characteristics of neurons known for rep-resenting visual speed. Specifically, we found that the population coding accuracy for visual speed in area MT (“neural prior”) is precisely predicted by the power-law, slow-speed prior extracted from fitting the Bayesian model to psychophysical data (“behavioral prior”) to the point that the two priors are indistinguishable in a cross-validation model comparison. Our results demonstrate a quantitative validation of the Bayesian observer model constrained by efficient coding at both the behavioral and neural levels. Significance Statement: Statistical regularities of the environment play an important role in shaping both neural representations and perceptual behavior. Most previous works addressed these two aspects independently. Here we present a quantitative validation of a theoretical framework that makes joint predictions for neural coding and behavior, based on the assumption that neural representations of sensory information are efficient but also optimally used in generating a percept. Specifically, we demonstrate that the neural tuning characteristics for visual speed in brain area MT are precisely predicted by the statistical prior expectations extracted from psychophysical data. As such, our results provide a normative link between perceptual behavior and the neural representation of sensory information in the brain.
Open AccessPrior expectations in visual speed perception predict encoding characteristics of neurons in area MT
Bayesian inference provides an elegant theoretical framework for understanding the characteristic biases and discrimination thresholds in visual speed perception. However, the framework is difficult to validate due to its flexibility and the fact that suitable constraints on the structure of the sensory uncertainty have been missing. Here, we demonstrate that a Bayesian observer model constrained by efficient coding not only well explains human visual speed perception but also provides an accurate quantitative account of the tuning characteristics of neurons known for rep-resenting visual speed. Specifically, we found that the population coding accuracy for visual speed in area MT (“neural prior”) is precisely predicted by the power-law, slow-speed prior extracted from fitting the Bayesian model to psychophysical data (“behavioral prior”) to the point that the two priors are indistinguishable in a cross-validation model comparison. Our results demonstrate a quantitative validation of the Bayesian observer model constrained by efficient coding at both the behavioral and neural levels. Significance Statement: Statistical regularities of the environment play an important role in shaping both neural representations and perceptual behavior. Most previous works addressed these two aspects independently. Here we present a quantitative validation of a theoretical framework that makes joint predictions for neural coding and behavior, based on the assumption that neural representations of sensory information are efficient but also optimally used in generating a percept. Specifically, we demonstrate that the neural tuning characteristics for visual speed in brain area MT are precisely predicted by the statistical prior expectations extracted from psychophysical data. As such, our results provide a normative link between perceptual behavior and the neural representation of sensory information in the brain.
- Open Access
An image reconstruction framework for characterizing initial visual encoding
We developed an image-computable observer model of the initial visual encoding that operates on natural image input, based on the framework of Bayesian image reconstruction from the excitations of the retinal cone mosaic. Our model extends previous work on ideal observer analysis and evaluation of performance beyond psychophysical discrimination, takes into account the statistical regularities of the visual environment, and provides a unifying framework for answering a wide range of questions regarding the visual front end. Using the error in the reconstructions as a metric, we analyzed variations of the number of different photoreceptor types on human retina as an optimal design problem. In addition, the reconstructions allow both visualization and quantification of information loss due to physiological optics and cone mosaic sampling, and how these vary with eccentricity. Furthermore, in simulations of color deficiencies and interferometric experiments, we found that the reconstructed images provide a reasonable proxy for modeling subjects’ percepts. Lastly, we used the reconstruction-based observer for the analysis of psychophysical threshold, and found notable interactions between spatial frequency and chromatic direction in the resulting spatial contrast sensitivity function. Our method is widely applicable to experiments and applications in which the initial visual encoding plays an important role.
Open AccessAn image reconstruction framework for characterizing initial visual encoding
We developed an image-computable observer model of the initial visual encoding that operates on natural image input, based on the framework of Bayesian image reconstruction from the excitations of the retinal cone mosaic. Our model extends previous work on ideal observer analysis and evaluation of performance beyond psychophysical discrimination, takes into account the statistical regularities of the visual environment, and provides a unifying framework for answering a wide range of questions regarding the visual front end. Using the error in the reconstructions as a metric, we analyzed variations of the number of different photoreceptor types on human retina as an optimal design problem. In addition, the reconstructions allow both visualization and quantification of information loss due to physiological optics and cone mosaic sampling, and how these vary with eccentricity. Furthermore, in simulations of color deficiencies and interferometric experiments, we found that the reconstructed images provide a reasonable proxy for modeling subjects’ percepts. Lastly, we used the reconstruction-based observer for the analysis of psychophysical threshold, and found notable interactions between spatial frequency and chromatic direction in the resulting spatial contrast sensitivity function. Our method is widely applicable to experiments and applications in which the initial visual encoding plays an important role.