Reimagining the binding problem(s) for the 21st century

Symposium: Friday, May 15, 2026, 10:30 am – 12:30 pm, Talk Room 2

Schedule of Events | Search Abstracts | Symposia | Talk Sessions | Poster Sessions

Organizers: JohnMark Taylor¹, Seda Karakose-Akbiyik²; ¹Columbia University, ²Stanford University
Presenters: Peter Tse, JohnMark Taylor, Seda Karakose-Akbiyik, Ana Chica, Anne Sereno, Jake Quilty-Dunn

Research on the binding problem stands at an exciting juncture: decades of psychophysical research have illuminated the dynamics of feature binding, increasingly detailed neural measurements have mapped how features are represented across early, ventral, and dorsal visual regions, and deep neural networks (DNNs) have emerged as powerful “model organisms” for testing candidate models of visual processing, as well as valuable tools for real-world tasks requiring successful binding. However, this wealth of research spread across studies employing different paradigms or studying different parts of the brain—and often studied in separate research traditions—has given rise to the “methodological binding problem”: synthesizing these disparate findings into unified neurocognitive models of how the brain integrates visual features. Indeed, just as the presence of multiple objects raises the challenge of associating them with the correct features, there may be multiple binding problems, and it is important to establish when different subfields are referring to the same one. This symposium tackles both the classical and methodological binding problems head-on by bringing into dialogue new developments from neurophysiology (spanning early, ventral, and dorsal visual regions), neuropsychology, psychophysics, computational modeling, and philosophy across six talks and a subsequent panel discussion. With the goal of renovating the binding problem for the 21st century, all six talks either present evidence that challenges us to reconceptualize core components of the binding problem as traditionally posed, or leverage new methods to make progress on important avenues inaccessible to earlier researchers. One talk presents psychophysical and neural evidence that visual search for specific feature conjunctions can approach pop-out with training, challenging the traditional view that only elementary features, not conjunctions, can be encoded pre-attentively. Another talk presents neuroimaging evidence against the neuroanatomical segregation of different features, such as color and shape, that has often been used to motivate the existence of a binding problem in the first place, and compares how the human brain and DNNs integrate visual features. New neuropsychological findings illuminate how the ventral and dorsal streams may interact during feature integration, a crucial but underexplored piece of the puzzle. Another line of work measures brain activity using fMRI and EEG during stimulus conditions designed to induce binding errors, identifying oscillatory signatures and occipito-parietal activation patterns that distinguish correct and incorrect binding (overcoming the longstanding methodological challenge of identifying neural signatures reflecting conscious binding rather than the mere co-occurrence of features). Another talk uses DNNs to identify key computational benefits of segregated feature processing (illuminating why the binding problem may arise to begin with), and provide powerful new tests of longstanding hypotheses about how space constrains feature binding. Finally, a talk from philosophy argues for a novel taxonomy of multiple forms of binding, developing much-needed conceptual clarity and setting the stage for a concluding panel discussion that seeks to relate different findings, establish theoretical common ground among research traditions, and articulate remaining gaps in our understanding. The symposium aims at a broad audience, reflecting its interdisciplinary nature and goal of “binding” insights across different subcommunities in VSS.

Talk 1

Learning to master our world: conjunction learning in the brain and mind

Peter Tse¹, Sebastian Frank², Eric Reavis³, Mark Greenlee²; ¹Dartmouth University, ²University of Regensburg, ³UCLA

We do not come to the world recognizing the conjunction of redness and the octogonal shape that together comprise a stop sign. Yet, through exposure and with practice, we automatize this conjunction into a unit that speeds processing. Such conjunction learning happens both motorically and perceptually. Serial search for a conjunction-defined target typically involves slower reaction times in visual search the more distractors there are. We find that with practice, search slopes flatten such that conjunction search comes to resemble visual pop-out. This learning is not like traditional forms of perceptual learning in that it transfers to locations, sizes and orientations and even to other conjunctions that differ from those used in training. In this talk I will ask what actually is learned when a conjunction is learned, and how such learning happens in the brain, both cortically and subcortically.

Talk 2

Joint population coding of color and form in the human ventral pathway and convolutional neural networks

JohnMark Taylor¹, Yaoda Xu²; ¹Columbia University, ²Yale University

Both the human ventral visual pathway and CNNs trained to recognize objects have been shown to encode color and form information about stimuli, but how exactly are these features jointly represented in distributed population codes, and how might the nature of this representation vary between biological and artificial visual systems? This talk examines population coding of color and form across multiple experiments, using both controlled artificial stimuli and naturalistic stimuli. In the human ventral pathway, color and form information were found to be decodable using fMRI in every region in the ventral pathway for artificial stimuli, even in regions defined based on their univariate sensitivity to one feature or the other, with the exception of a color-sensitive region anterior to V4. However, despite this anatomical intermingling, color and form information tended to remain representationally independent, with brain regions encoding information about the two individual features, but generally not about particular conjunctions of features. By contrast, in CNNs, we found in a study using naturalistic colored stimuli that color and form tended to be encoded in an increasingly conjunctive format as processing proceeded. That said, further analyses suggest that the ventral pathway and CNNs trained to recognize objects are similar in their relative emphasis of these two features, with both systems increasingly emphasizing color relative to orientation information, but increasingly deemphasizing color relative to curvature information. Collectively, this work thus identifies both similarities and divergences in how these systems solve the problem of simultaneously encoding multiple features across distributed neural populations.

Talk 3

Insights about feature binding from neuropsychological deficits of color and shape processing

Seda Karakose-Akbiyik¹, Alfonso Caramazza²; ¹Stanford University, ²Harvard University

Color and shape are classically thought to be processed via partially segregated cognitive and neural mechanisms, giving rise to a binding problem for generating our unitary perceptual experience. Spatial attention is suggested to play a role in integrating visual features, but the mechanisms underlying this process are not fully understood. In addition, recent neurophysiological evidence on joint encoding of visual features in the ventral visual stream has called into question the existence of a binding problem. One way to get a handle on this debate is to study deficits of visual perception that arise after brain damage. In this talk, we will review deficits in perceiving visual features and their conjunctions that arise after brain damage, using color and shape processing as a case study. We will discuss the theoretical implications of these deficits for our accounts of visual feature encoding and binding. We will review work on visual agnosia, achromatopsia, and attentional failures as well as our recent findings from a patient with neglect (DH) who is able to report color, orientation, and location of shapes on the left half of objects despite lack of awareness of the shapes themselves. We propose that previous neuropsychological evidence on color and shape processing and DH’s novel behavioral profile provide intriguing insights regarding visual feature encoding and integration, interactions between dorsal and ventral visual regions, and the mechanisms of spatial attention in this process.

Talk 4

Brain oscillations and occipito-parietal contributions to color-shape integration

Ana Chica¹, Maria Cobos¹, Pablo Rodríguez-San Esteban¹; ¹University of Grenada

Our sensory system constantly receives information from both the environment and our own bodies. Despite our impression to the contrary, we are largely unaware of most of this information and often cannot report it accurately. Thus, perception—a seemingly ordinary process—can be complex and requires significant computational effort, which can lead to errors such as incorrect feature integration. We will present data from two studies that used functional magnetic resonance imaging (fMRI) and electroencephalography (EEG) to investigate the neural mechanisms associated with correct and incorrect feature integration. fMRI data showed that a distributed set of brain regions exhibited increased activity during correct feature integration compared to incorrect integration, including enhanced functional coupling between occipital, parietal and frontal regions. In contrast, incorrect feature integration was linked to heightened occipital (V1-V2) responses in the early stages of processing, along with a general decrease in coactivation between occipital, parietal and frontal regions. EEG data indicated that alpha and beta power were enhanced during correct as compared to incorrect feature integration. This result was associated with cognitive processes such as pre-stimulus preparation, distractor inhibition or the reliability of perceptual information representation within the cortex. Our findings underscore the role of parietal regions in feature integration and highlight the importance of occipital-parietal interactions, which oscillate in low-frequency bands, in perceptual processing.

Talk 5

Insights into the reasons for segregated processing and the privileged role of space in binding, from a computational perspective

Anne Sereno¹, Zhixian Han²; ¹Purdue University, ²Indiana University

Much neuroscience research supports segregation of visual processing. Using artificial neural networks (ANNs), we previously showed that segregated streams for identity and space actively retain information about both identity and space, independently and differently; and further, that ANNs with two pathways, compared to a single pathway, have better performance (higher accuracy, less training time); with these advantages increasing with binding difficulty (i.e., increasing number of objects), suggesting segregation itself may help ease the binding problem. However, if segregated-pathway networks have advantages over single-pathway networks, why are not all attributes segregated in the brain? Simulations suggest that when there are a greater number of classes or greater discrimination is needed for a given attribute, the advantage of segregated pathways is greater. These ANNs used a relative location map to successfully bind identity and absolute location of objects. However, the brain needs to combine many different attributes (e.g., color, identity) and any of these attributes could be used in a similar fashion to constrain the binding problem. Using visual objects with multiple attributes (identity, luminance, orientation, location) to be recognized, we tried to find the best attribute (among an identity, luminance, orientation or location map) to constrain the binding problem. When attributes are independent, a location map is always the best choice for constraining the binding problem. A computational approach can provide insights into why segregation and the binding problem arise and the best way to constrain the binding problem. Implications of the findings for neuroscience and modeling will be discussed.

Talk 6

Two ways to see a red square

Jake Quilty-Dunn¹; ¹Rutgers University

Vision scientists and philosophers share an interest in how representations are combined in visual perception. Much of the work on this topic presupposes that features are bound by being attributed to the same object. However, this “object-based” account of binding fails to explain cases of binding that occur outside attentive, object-centric processing. The philosophical arguments for the object-based account also fail to establish that attributing features to objects is necessary for binding. This talk explores a range of findings and argues that the visual system has multiple different principles for binding features. Object-based binding is especially useful for interfacing with systems outside vision, including verbal report. Other forms of binding are more useful for processes internal to the visual system, including ensemble coding. The resulting picture is that vision constructs redundant representations of feature conjunctions for different purposes.