Preamble

Interaction Between Dorsal and Ventral Pathways in Visual Object Representation Processing

School of Psychology, Shandong Normal University; School of Psychology, Nanjing Normal University, Nanjing.

The visual system of the brain processes visual information through the dorsal and ventral pathways. Traditionally, the ventral pathway is primarily responsible for the visual recognition of objects (the "what" system), while the dorsal pathway is responsible for processing the spatial location and motion of objects (the "where" system). However, the existence of multiple interactive neural connections between these pathways suggests a functional interaction during the processing of visual object representations.

The ventral pathway alone cannot fully account for the representation of an object's global shape. It is necessary for the dorsal pathway's representation of global shape information to converge with the local feature information represented by the ventral pathway. In the context of goal-directed thought and behavior, the dorsal pathway must extract and maintain visual information from the ventral pathway in real-time to achieve adaptive visual processing. The former represents a functional integration primarily driven by features, while the latter represents a functional integration primarily driven by goals.

Future research should focus on several key areas: investigating the influence of attention on the dorsal pathway's representation of global shape; exploring how object familiarity modulates the integration of global shapes and local features; and examining the mechanisms by which visual working memory resists interference to maintain the processing of goal-directed stimuli. Furthermore, it is essential to study the impact of endogenous memory information on adaptive visual object representation, as well as the developmental mechanisms of the dorsal and ventral pathways and their subsequent effects on pathway interaction.

关键词

Visual Object Representation and the Interactions within the Ventral Stream

Visual object representation is a fundamental process in the primate brain, primarily mediated by the ventral visual stream. This pathway, extending from the primary visual cortex (V1) to the inferior temporal (IT) cortex, is responsible for transforming raw retinal inputs into sophisticated neural representations that allow for robust object recognition. Central to this process is the complex interaction between different hierarchical levels of the ventral stream, which enables the brain to resolve the competing demands of specificity and generalization.

Invariant Processing of Visual Object Information

A hallmark of the ventral stream is its ability to achieve "object invariance." This refers to the capacity of the visual system to recognize an object despite significant changes in its retinal image caused by variations in position, scale, viewpoint, and illumination. As information progresses through the hierarchy—from V1 to V2, V4, and finally the IT cortex—the receptive fields of neurons increase in size, and their tuning properties become increasingly complex.

In the early stages, neurons respond to simple features like edge orientations. However, in the higher-level IT cortex, neurons exhibit selectivity for complex shapes and objects while maintaining consistent firing rates regardless of the object's physical transformation. This invariant representation is crucial for navigating a dynamic environment, ensuring that a "chair" is recognized as such whether it is viewed from the front, the side, or under different lighting conditions.

Adaptive Processing of Visual Object Information

While invariance provides stability, the visual system must also remain highly adaptive to meet the demands of specific tasks and environmental changes. Adaptive processing involves the dynamic modulation of neural responses based on context, attention, and prior experience. This flexibility allows the brain to transition from coarse categorization (e.g., identifying an animal) to fine-grained discrimination (e.g., identifying a specific breed of dog) depending on the observer's goals.

Research suggests that this adaptability is driven by both bottom-up sensory input and top-down signals from the prefrontal cortex and other high-level areas. These interactions allow the ventral stream to "re-tune" its representations. For instance, through perceptual learning, the visual system can become more sensitive to the subtle features that distinguish highly similar objects. This plasticity ensures that visual representations are not static templates but are instead dynamic constructs that can be optimized for behavioral relevance.

Interactions within the Ventral Pathway

The synergy between invariant and adaptive processing is made possible by the dense recurrent connections—both horizontal and feedback—

1 引言

The visual system of the brain processes objects through the dorsal and ventral pathways (Vinci-Booher et al., 2022). Both pathways receive input from the primary visual cortex (V1), diverging after passing through the V2 area. The ventral pathway projects through V4 to the inferior temporal cortex (IT), forming the "what" pathway responsible for object visual recognition and processing. In contrast, the dorsal pathway projects through V3 and V5/MT to the posterior parietal cortex (PPC), forming the "where" or "how" pathway that serves vision-guided actions (Freud et al., 2020; Freud et al., 2016; 2023; Mahon, 2023). Although substantial evidence supports the view that information processing in these two pathways is segregated, an increasing number of studies have revealed multiple interactive neural connections between them. This suggests that functional interactions between the dorsal and ventral pathways exist during the processing of visual object representations (Freud & Behrmann, 2020; Garcea et al., 2018; Goldstein-Marcusohn et al., 2024; Vinci-Booher et al., 2022; 2020).

Invariant and adaptive visual information processing are two key dimensions of how the brain represents visual objects (2018b). The former refers to the visual system's ability to achieve object recognition and categorical processing through invariant representations of object features, even when perceptual conditions (such as viewpoint or lighting) change (Ayzenberg & Behrmann, 2024; Nestmann et al., 2022; Vaziri-Pashkam et al., 2022). The latter refers to the visual system's capacity to achieve adaptive visual object representation aligned with an individual's goals and tasks by dynamically selecting and temporarily storing object information (2018a, 2018b, 2024). Traditional views held that invariant object processing was completed independently by the ventral pathway (Kravitz et al., 2013). However, recent research has found that the ventral pathway needs to integrate global shape information represented by the dorsal pathway to achieve object invariance (Ayzenberg & Behrmann, 2022a, 2023b, 2024). Furthermore, the dorsal pathway can extract and temporarily store object information from the ventral pathway in real-time via visual working memory (VWM) to facilitate adaptive visual information representation (Taylor et al., 2022; Vaziri-Pashkam et al., 2019; 2018a, 2018b, 2023a). Evidently, elucidating the interaction mechanisms between the dorsal and ventral pathways in visual object representation not only refines our understanding of the brain mechanisms underlying object cognition but also holds significant value for artificial intelligence modeling of object recognition and clinical research into visual perception disorders. To this end, this paper first introduces the structural composition and fiber connections of the dorsal and ventral pathways to clarify the neural basis for information exchange between them. It then analyzes the functional segregation and integration of these pathways in the invariant processing of visual object representations. Subsequently, it discusses the dynamic integration mechanism by which the dorsal pathway extracts and represents object information from the ventral pathway during goal-directed adaptive visual processing. Finally, the paper explores several key questions and mechanisms that require further in-depth research.

2 背

Structural Composition and Fibrous Connectivity of the Ventral Pathway

The structural composition and fibrous connectivity of the ventral pathway are illustrated in [FIGURE:1], where [FIGURE:1B] represents the macaque brain. In the dorsal pathway of the human brain, visual information is transmitted forward from areas V3A/B to the posterior parietal cortex (PPC). The PPC contains a continuous band composed of multiple regions distributed within and around the intraparietal sulcus (IPS) \cite{Kastner2017, Kravitz2013}. From posterior to anterior, these regions are sequentially identified as IPS0, IPS1, IPS2, IPS3, IPS4, and IPS5. Similarly, in the macaque brain, visual information in the dorsal pathway progresses forward through the following areas: the caudal intraparietal area (CIP), lateral intraparietal area (LIP), ventral intraparietal area (VIP), medial intraparietal area (MIP), and anterior intraparietal area (AIP) \cite{Freud2016, Kastner2017}. Within the ventral pathway of the human brain, visual information is transmitted via V4 to the lateral occipital cortex (LOC) at the occipitotemporal junction, and subsequently moves forward to the ventral occipitotemporal cortex (VOTC).

The VOTC includes the fusiform gyrus (FG), the parahippocampal gyrus (PHG), and the anterior temporal lobe (ATL) \cite{Ayzenberg2022a, Kravitz2013}. In the macaque brain, visual information in the ventral pathway travels from area V4 forward to the inferior temporal (IT) cortex \cite{Kravitz2013}. Evidence from anatomical studies and non-invasive diffusion magnetic resonance imaging (dMRI) indicates the existence of extensive and dense neural connections between the dorsal and ventral pathways \cite{Takemura2019}. Specifically, the vertical occipital fasciculus (VOF) connects the ventral occipital lobe with the dorsal visual areas (V3A/B, IPS0) \cite{Takemura2019}. Furthermore, the posterior vertical pathway (PVP) links broad regions of the dorsal and ventral pathways from the posterior to the anterior sections \cite{Bullock2019, Vinci-Booher2022}. In the macaque brain, these extensive regions between the dorsal and ventral pathways are connected by the inferior longitudinal fasciculus (ILF) and the middle longitudinal fasciculus (MdLF) \cite{Takemura2024, Roumazeilles2020}. It is evident that multiple interactive neural connections exist between the dorsal and ventral pathways, providing an anatomical foundation for the functional interactions between these pathways during the processing of visual object representations.

3 视觉

Functional Dissociation and Integration of the Dorsal and Ventral Pathways in Invariant Representation Processing

Recent research on non-human primates and human subjects (Ayzenberg & Lourenco, 2022; Ayzenberg & Simmons, 2023) indicates a clear functional dissociation between the dorsal and ventral pathways during object representation. Specifically, the dorsal pathway is primarily responsible for representing the spatial arrangement and structural configuration of objects.

[FIGURE:1]

Consider the example of four different aircraft. Although these four planes possess entirely distinct local features (such as variations in wing shape, engine placement, or fuselage texture), they share an identical underlying spatial arrangement. This common structural configuration allows the visual system to generate a consistent perception of the overall global shape, leading us to categorize all of them as "airplanes." This phenomenon underscores how the brain achieves invariant representation by prioritizing global spatial relationships over idiosyncratic local details.

Source:

Ayzenberg Behrmann, 2022a

The global shape of an object represents its overall structure, whereas local features characterize its specific details. Because global shape representation is object-centered, it describes the overall configuration of an object through the spatial arrangement of its parts while maintaining a degree of tolerance toward variations in the local features themselves \cite{Ayzenberg Behrmann, 2022a}. This characteristic makes global shape representation a critical organizational feature for supporting the processing of visual object invariance. Consequently, it is necessary for the ventral pathway to integrate global shape information represented by the dorsal pathway to achieve robust invariant processing of visual objects.

3.1 视觉物体表征加工中背

Functional Dissociation of the Ventral Pathway

Substantial evidence supports a functional dissociation between the dorsal and ventral pathways in the processing of visual object representations. First, the ventral pathway exhibits higher sensitivity to local features than to global shape. When subtle, imperceptible local perturbations are applied to images (e.g., identity-preserving distortions), single neurons in the monkey ventral stream exhibit strong and dense firing responses \cite{JagadeeshGardner2022}. These neurons are highly sensitive to changes in local features, even when such changes do not alter the perceived identity of the object. Furthermore, for synthetic images where the spatial arrangement of local features is scrambled, multivariate responses from human functional magnetic resonance imaging (fMRI) and the outputs of computational models based on monkey IT neurons fail to distinguish these scrambled images from intact natural images \cite{JagadeeshGardner2022}. This suggests that these two categories of images are represented similarly within the ventral cortex, despite the fact that human observers can easily differentiate between them. At both the single-neuron and population levels, the ventral pathway does not represent objects as holistic shapes but rather as collections of features, where the precise spatial arrangement of these features does not significantly influence the representation. For instance, "Texforms"—images that preserve certain texture statistics while disrupting clear contours and explicit shapes to the point of unrecognizability—fail to be identified by human observers. However, they activate the same large-scale functional maps in the ventral pathway as intact natural images, such as those representing object size and animacy \cite{Long2018, Long2022}. These findings indicate that the functional activation of large-scale cortical areas in the ventral pathway is better described by local features rather than global shape.

Independent Representation of Global Shape by the Dorsal Pathway

In contrast to the ventral pathway, the dorsal pathway can independently represent global shape. Unlike its response to synthetic images with scrambled local features, the dorsal pathway exhibits a significantly greater response to intact object images \cite{FreudCulham2017}. Research involving transcranial magnetic stimulation (TMS) and brain lesions has demonstrated that temporary inactivation of or damage to the dorsal cortex impairs the perception of global object shape, including configurational processing \cite{Zachariou2017}, the relationships between object features \cite{Thomas2012}, and visuospatial perception \cite{Medina2020}. Although the ventral cortex is the core region for object recognition, the dorsal cortex retains the ability to represent the three-dimensional (3D) structure of objects even after ventral damage \cite{FreudGanel2017}. \cite{Riddoch2008} identified a double dissociation between lesion location and functional deficits: patients with lesions restricted to the ventral cortex showed deficits in perceiving local features but retained global shape perception, whereas patients with lesions restricted to the dorsal cortex showed deficits in global shape perception but retained local feature perception. With advancements in computational vision, a quantitative model known as the "shape skeleton" has been developed to describe the spatial arrangement of an object's parts through a series of internal axes of symmetry \cite{Ayzenberg2022, AyzenbergLourenco2022}.

The shape skeleton model effectively explains how the human visual system represents global shape for object recognition \cite{Ayzenberg2019, AyzenbergLourenco2019, Destler2019, Wilder2019}. Compared to other visual models, the shape skeleton better predicts response patterns in the dorsal pathway and independently accounts for a significant portion of its response variance \cite{AyzenbergBehrmann2022b, Ayzenberg2022}. Thus, while the dorsal pathway primarily represents the global shape of visual objects, the ventral pathway is sensitive to local features, and both pathways contribute to object representation. Crucially, how do the dorsal and ventral pathways interact to achieve invariant processing of visual objects? Given that the ventral pathway is the central hub for object recognition and that invariant representation is a hallmark of ventral stream processing, Ayzenberg and Behrmann \cite{AyzenbergBehrmann2022a, AyzenbergBehrmann2024, AyzenbergSimmons2023} proposed that the dorsal pathway independently completes the representation of global shape at an early stage. This information is then transmitted to the ventral pathway to support the invariant representation and processing of visual objects.

3.2 视觉物体不变性表征加工中背

The integration mechanism of the ventral pathway serves as a neural locus for functional integration and is a critical region for object recognition. Along the posterior-anterior axis, it is divided into two subregions: the lateral occipital cortex (LO) and the posterior fusiform gyrus (pFs) \cite{Grill-Spector1999}. These areas exhibit the highest shape sensitivity within the ventral pathway, and their representational similarity to the posterior regions of the dorsal pathway (V3a-IPS0) is greater than that of other ventral regions \cite{Freud2017, Culham2017}. Even when representing complex objects, the similarity between these regions and the posterior dorsal pathway remains higher than elsewhere in the ventral stream \cite{Ayzenberg2023}, potentially reflecting a shared representation of holistic facial configurations \cite{Zachariou2017}. Thus, although these two regions are not located within the same pathway, they possess more similar shape representation structures, providing a neural basis for receiving holistic shape information projections from the dorsal pathway \cite{Ayzenberg2023, Simmons2023}. In fact, as part of the ventral pathway, LO and pFs also follow a hierarchical processing structure along the posterior-anterior axis \cite{Lerner2002, Caramazza2022}. LO is more sensitive to local features of objects, whereas pFs is more sensitive to the global integrity of objects and exhibits more stable invariant representations across changes in object size and position. Furthermore, pFs can decode the skeletal information of object shapes, whereas no evidence for such skeletal representation has been found in LO \cite{Lescroart2013, Biederman2013}. This reflects a progression in information integration where pFs may primarily be sensitive to semantic-level information \cite{Ayzenberg2022}. Taken as a whole, LO and pFs facilitate the functional integration of the ventral pathway. Specifically, upon receiving rapid projections of holistic shape information from the dorsal pathway, LO and pFs integrate this with local features as the processing hierarchy ascends, ultimately achieving invariant processing of visual objects \cite{Ayzenberg2023, Simmons2023, Nestmann2021}. This process is supported by the fact that visual information processing in the dorsal pathway occurs earlier than in the ventral pathway.

Electrophysiological studies in humans and non-human primates have revealed a "latency advantage" in the visual information processing of the dorsal pathway. For instance, single-cell recordings in monkeys indicate that shape selectivity in the dorsal pathway emerges approximately 30–40 ms after stimulus onset, whereas the latency in the ventral pathway is approximately 60–70 ms \cite{Janssen2008}. Similarly, for shape selectivity, the latency in the dorsal pathway is approximately 80 ms, compared to approximately 110 ms in the ventral pathway \cite{Theys2012}. Human electroencephalography (EEG) \cite{Regev2018} and magnetoencephalography (MEG) \cite{Ayzenberg2017} studies have also found that the decoding time for holistic shapes in the dorsal pathway (with latencies of 60–90 ms) is earlier than in the ventral pathway (with latencies of 100–130 ms). The temporal advantage of the dorsal pathway in visual processing may stem from its reception of coarse, low-contrast, and low-spatial-frequency object information rapidly transmitted via the magnocellular pathway \cite{Ayzenberg2023, Simmons2023}. In contrast, the ventral pathway receives fine, high-contrast, and high-spatial-frequency information transmitted more slowly via the parvocellular pathway \cite{Collins2019, Collins2023}. Sensitivity to coarse object information in the dorsal pathway appears as early as 50 ms post-stimulus, while the ventral pathway's sensitivity to fine information emerges at the earliest around 100 ms \cite{Collins2019}. Crucially, the coarse information transmitted through the magnocellular pathway is sufficient to compute the holistic shape of an object \cite{Ayzenberg2023, Simmons2023}. The dorsal pathway transmits these holistic shape representations to the ventral pathway to support invariant object representation processing. Ayzenberg and Behrmann (2022b) directly investigated the influence of the dorsal pathway on the ventral pathway during holistic shape representation. They found that the posterior intraparietal sulcus (pIPS) computes part-relations of objects and exhibits object category decoding capabilities comparable to LO. Importantly, mediation and multivariate effective connectivity analyses demonstrated that the multivariate responses of pIPS mediate the holistic shape representations in LO, showing significant effective connectivity where pIPS processing precedes and predicts LO processing. However, the low temporal resolution of fMRI precludes definitive conclusions regarding the precise time course and directionality of these cortical interactions. To address this, Ayzenberg and Simmons (2023) utilized high-density EEG—a technique with high temporal precision and spatial resolution—to explore this issue. They found that the decoding time for information in the dorsal pathway (66 ms) was significantly earlier than in the ventral pathway (94 ms), and that the dorsal pathway predicts ventral pathway responses in a time-dependent manner.

Functional dissociation exists within the ventral pathway during visual object representation, necessitating the integration of holistic shape information from the dorsal pathway for invariant processing. Coarse object information from the primary visual cortex is rapidly transmitted to the dorsal pathway via the magnocellular pathway, forming a holistic shape representation during the early stages of processing. This representation is then integrated with fine local feature information arriving via the parvocellular pathway, subsequently progressing anteriorly along the ventral pathway.

This integration completes the invariant representation processing of visual objects. This process constitutes a mode of object information processing that is primarily driven by features.

4 视觉物体的适应性表征加工中背

Functional Integration of the Ventral Pathway

In the early stages of visual object processing, representation of global shape information within the dorsal pathway can provide critical support for the invariant representation processing occurring in the ventral pathway. The ventral pathway is not only capable of rapid object recognition but also provides a stable and detailed analysis of the visual environment. However, if the visual system were to indiscriminately make all visual input available, it would likely lead to significant distraction and disrupt task execution. At any given moment, the visual system must exclude the vast majority of task-irrelevant information, selecting only a small subset of relevant data to support cognitive processes such as task planning and problem-solving. This selection process guides appropriate behavior to complete goal-directed tasks \cite{2018a, 2018b, 2023a}.

Consider the example of observing a cat wearing a red bow. An invariant visual system would faithfully reflect the cat's visual features, maintaining consistency across general characteristics, contexts, and tasks. In contrast, an adaptive visual system emphasizes different visual features of the cat based on the individual's focus of attention. For instance, if an individual wants to identify the type of animal, the system will emphasize the cat's conceptual attributes; however, if the goal is to determine the color of the bow, the system will prioritize that specific feature \cite{2018b}.

The adaptive representation processing of visual objects is a goal-directed process based on the current task. It requires a functional integration platform capable of dynamically selecting and representing the visual information processed within the ventral pathway. In this context, the dorsal pathway likely plays a crucial role in facilitating this integration.

4.1 PPC

The dorsal pathway for visual objects serves as an ideal cortical region for the processing of adaptive visual representations \cite{2018a, 2018b, 2020}. A series of functional regions are organized along a posterior-to-anterior gradient. Located posteriorly are the homologous regions in monkeys responsible for spatial information processing \cite{Alizadeh 2018; Medina 2020}. These include the human parietal grasp region (AIP) and the human parietal reach region (MIP) located in the medial portion of the intraparietal sulcus, which are primarily involved in motor execution such as eye movements, grasping, and reaching \cite{Kastner 2017}. The regions IPS1/IPS2 (LIP) and the superior intraparietal sulcus (sIPS) \cite{Bettencourt 2016b, 2018a} are primarily responsible for the cognitive processing of visual object information \cite{2018a, 2018b, 2020} and participate in the storage and encoding of visual object features \cite{Lefco 2020}. Furthermore, the bilateral superior parietal lobule (SPL) and the right temporal parietal junction (TPJ) are mainly responsible for attentional control and flexible shifting \cite{2024; 2018a}. Consequently, the LIP, which is responsible for cognitive processing, is surrounded by regions that handle visuospatial information, motor actions, and attentional control. Through attentional control mechanisms, this brain region can determine which visual information is selected and processed, and subsequently store and flexibly represent it to meet the demands of adaptive task processing.

The LIP supports the adaptive representational processing of visual objects, which requires two key characteristics \cite{2018a, 2018b}. First, it must possess strong attentional control and task-modulation capabilities to focus attention on the target task while resisting interference from distractors. Second, it must be capable of storing and representing task-relevant visual information. When task-relevant visual information is no longer visible, the visual processing system must maintain this information for a period to facilitate further cognitive processing. Visual processing in the sIPS (LIP) is modulated by attention and task demands \cite{Taylor 2024; Vaziri-Pashkam 2019}. For instance, many neurons in the monkey LIP exhibit selectivity for visual features only when they are task-relevant \cite{Seideman 2022}. Multi-voxel pattern analysis (MVPA) in humans has also demonstrated that visual features such as shape \cite{Jeong 2015}, color \cite{Shim 2017}, and category \cite{Bracci 2017} can only be successfully decoded in the LIP when they are task-relevant or attended to, whereas the region does not encode task-irrelevant shapes \cite{2010}. The LIP also demonstrates robust suppression of distractors \cite{Vaziri-Pashkam 2017}; its object classification accuracy remains stable under distractor interference and correlates with individual behavioral performance \cite{Bettencourt 2016a; 2024}. Patients with damage to this area are unable to perform appropriate hand movements based on task requirements \cite{Goodale, Meenan, 1994}. When competitive distractors are present, patients with reversible lesions in this region show a sharp decline in the correct identification of cued target stimuli during the early stages of injury; however, after recovery (e.g., 40 days post-injury), their resistance to competitive distractors does not differ from normal controls \cite{Gillebert 2011}. Task demands and stimulus features jointly modulate the encoding of visual information by LIP neurons \cite{Ibos Freedman, 2016} and the strength of visual representations \cite{Long Kuhl, 2018; Vaziri-Pashkam 2017}. Taylor and Xu \cite{2024} utilized multidimensional scaling (MDS) to analyze representational dissimilarity matrices (RDMs) containing all combinations of tasks and categories. They found that within-task category representational geometries were more similar than between-task geometries, and that category representational geometries shared similar distributions across different tasks. In the LIP, representations of objects from the same category are more similar, and the distribution of category representations is not modulated by the task. This suggests that category and task jointly shape the representational geometry of the LIP, whereas the representational geometry of the ventral pathway is primarily dominated by object category. This indicates that the LIP not only selects information based on task goals but also further represents it. The sIPS (LIP) is also involved in storage and representation \cite{Kastner 2017; Lefco 2020}. Neurophysiological studies in monkeys have shown that the LIP persistently represents shape during delay periods \cite{Fitzgerald 2011; Sereno Maunsell, 1998}. Human research also indicates that LIP activity during working memory tasks is positively correlated with memory capacity \cite{Bettencourt 2016b; Sheremata 2018}. Importantly, during the delay period, the presence or predictability of distractors does not affect the decoding ability of the LIP or individual behavioral performance, whereas it significantly impacts the decoding ability of the early visual cortex \cite{Bettencourt 2020}.

This suggests that when distractors are present, the LIP maintains more robust representations than the early visual cortex, ensuring that information maintenance is not compromised by interference. Studies of brain damage \cite{Berryhill Olson, 2008} and transcranial magnetic stimulation \cite{Tseng 2010} have also found that lesions or interference in this region impair the storage and encoding of visual information. The LIP is essential for storage and representation; its close correlation with behavioral performance and its resistance to interference make it particularly suitable for the adaptive representational processing of visual objects. The sIPS (LIP) serves as a high-level cognitive hub related to attention and executive function \cite{Kastner 2017}. This region can select and temporarily maintain task-relevant visual information to support adaptive processing. While the invariant object processing of the ventral pathway requires support from the global shape representations of the dorsal pathway, the task-related adaptive visual processing of the dorsal pathway conversely requires object information represented by the ventral pathway \cite{2018a, 2018b}.

4.3 背

Extraction and Utilization of Object Visual Information from the Ventral Pathway by the Lateral Pathway. Human neurofunctional imaging (Vaziri-Pashkam & Kanwisher, 2019) and monkey neurophysiological studies (Borra & Luppino, 2017; Theys et al., 2015) have demonstrated that the representation of visual object information in the dorsal pathway is highly robust. This information includes simple features such as color, size, and texture, as well as complex features such as identity and object category (Freedman & Ibos, 2018), which are independent of spatial and motion information (Vaziri-Pashkam & Kanwisher, 2018a, 2018b). Numerous studies indicate that the visual object information represented in the dorsal pathway originates from the ventral pathway. First, the ventral pathway provides invariant object information to support the completion of goal-directed tasks. As previously mentioned, the lateral occipital complex (LO) is not only a key brain region in the ventral pathway for integrating global shape and local features but also exhibits invariance to object size and viewpoint (Nestmann et al., 2022; Vaziri-Pashkam & Kanwisher, 2022). IPS1–2 also exhibits object feature invariance similar to that of the LO, suggesting that integrated visual information from the LO may be utilized to support adaptive visual information processing in IPS1–2 (Konen & Kastner, 2008; Vaziri-Pashkam & Kanwisher, 2019). Neuropsychological research has shown that damage to the LO impairs the processing of invariant visual object information (Milner, 2017), as seen in patient D.F.

The visual information guiding the wrist rotation of patient D.F. appears to be limited to a single dimension (either the vertical or horizontal axis of a T-shaped aperture) (Goodale et al., 1994). Due to the damage in the LO, the patient is unable to integrate information regarding the two different axes of the T-shaped aperture; instead, they can only extract local features and orientation information from a single axis. For instance, when the horizontal axis of the T-shaped aperture is the target orientation, the patient consistently attempts to align the vertical axis of the T-shaped object with the horizontal axis of the aperture, leading to task failure. This demonstrates that the invariant object information represented by the LO is crucial for supporting the completion of goal-directed tasks (Vaziri-Pashkam & Kanwisher, 2018a, 2018b). Furthermore, structural connections and information flow exist between the LO and the IPS. Jitsuishi and Yamaguchi (2020) identified a specialized neural fiber bundle, the IPS-FG, which directly connects these two regions.

Multiple subregions of IPS1–2 are closely connected, and these subregions participate in cognitive processing at various levels, such as color analysis and object recognition. Consequently, the IPS-FG bundle can satisfy various cognitive functions, such as the requirement for integrating information from the ventral pathway. For example, research has found that viewing images of tools increases the functional connectivity between the LO and the IPS (Chen et al., 2018). Dynamic causal modeling (DCM) analysis of the effective connectivity between the dorsal and ventral cortex in the representation of tool knowledge (e.g., an axe) indicates that functional knowledge tasks enhance the connection strength between the primary visual cortex and the LO. In contrast, manipulative knowledge tasks enhance the connection strength from the primary visual cortex to the IPS (Kleineberg & Frey, 2018). This suggests that more abstract knowledge regarding object identity and function is first decoded in the ventral pathway, after which specific information is selectively transmitted to the dorsal pathway for further representation based on task requirements, ultimately completing visually guided behavior (Mahon, 2023). The dorsal pathway not only top-down selects and extracts visual information from the ventral pathway to support goal-directed behavior but also monitors task execution in real-time. Based on feedback signals, the dorsal pathway can top-down dynamically regulate and reshape object representations in the ventral pathway to improve the availability of object information and enhance recognition capabilities (Budisavljevic et al., 2018). Using representational dissimilarity matrix (RDM) analysis to examine representational similarity during the maintenance phase (Vaziri-Pashkam et al., 2023b), researchers found that representations in the LO during the maintenance phase were more similar to the representations in the IPS during the same phase than to the LO's own representations during the encoding phase.

Further analysis indicates that the LO automatically encodes a large amount of perceptual information about objects, at which point the LO is primarily driven by visual features. When specific object information is encoded and retained in the IPS according to task demands, the LO closely follows the task-relevant features held in the IPS.

In addition to directly acquiring object information, the dorsal pathway can top-down regulate and reshape the content related to the LO to better achieve cross-pathway integration of visual representations. The anatomical structure and functional characteristics of the IPS make it suitable as an integration platform, allowing it to dynamically select visual object information processed in the ventral pathway to adapt to the task requirements of visual representation processing. The sIPS (LIP) ensures that task-relevant information is temporarily stored and represented, allowing the content of visual representations to be controlled by attention to a greater extent and preventing distraction. This achieves the dynamic integration of adaptive object information—a process that is primarily a task-driven (top-down) form of visual object information processing.

5 总结与展望

The division of the brain's visual system into anatomically and functionally distinct dorsal and ventral streams is one of the most influential theoretical frameworks in cognitive neuroscience. However, the presence of multiple interactive neural connections between the dorsal and ventral streams suggests that there is a significant degree of integration and crosstalk between these two pathways.

Functional Interactions between Visual Pathways

The ventral pathway does not characterize global shapes—which are essential for invariant visual object processing—in isolation. Instead, it must receive global shape representation information transmitted from the posterior intraparietal sulcus (pIPS/IPS0). This information is then integrated with the local features represented within the ventral pathway to support invariant visual object processing. Consequently, the ventral pathway serves as a platform for the dynamic integration of object information, allowing for the flexible selection of appropriate visual data from the ventral stream based on specific task requirements.

In the representations and processing within the superior intraparietal sulcus (sIPS/IPS1–2), goal-directed adaptive visual information processing is achieved. This demonstrates a clear functional interaction between the dorsal and ventral pathways during visual object representation. Specifically, the former represents a primarily feature-driven (bottom-up) functional integration that provides us with stable, realistic, and detailed representations of the visual environment. In contrast, the latter represents a primarily task-driven (top-down) functional integration that enables flexible and effective communication and interaction with the external world. While the functional interactions between these pathways in visual object representation have been initially elucidated, several key questions and underlying mechanisms still require in-depth investigation.

The Influence of Global Shape

Xu (2023a) attempted to explain the global shape representation in the dorsal pathway from the perspective of attentional processing, suggesting that global shape is merely a byproduct of attentional processing rather than an inherent property of dorsal object representation. Attentional processes during object perception can be divided into bottom-up processing driven by features and top-down processing driven by cognitive control \cite{KatsukiConstantinidis2014}. The former is involved in the automatic integration of features such as color and size during early stages (75–100 ms), while the latter allocates attentional resources according to task goals during later stages (after 200 ms) \cite{Conci2011, Schneider2012}. However, the decoding time for an object's global shape \cite{AyzenbergSimmons2023} is earlier than the time required for feature integration \cite{Conci2011}, suggesting that the representation of global shape may be completed earlier.

Nevertheless, the mechanism of how attention participates in the dorsal pathway's representation of global shape remains unclear. During the process of selecting a target from competing distractor stimuli, attention prioritizes the global shape of an object rather than its local features \cite{White2021}. That is to say, the saliency of the global shape automatically attracts attentional resources and receives preferential representation. It should be noted that the medial axis (shape skeleton), which describes the structural information of an object's global shape, highlights global shape features \cite{AyzenbergLourenco2019, 2022}. The dorsal pathway primarily processes information transmitted via the magnocellular pathway, which is sensitive to low spatial frequencies \cite{FellemanVanEssen1991}, and participates in representing salient features in visual scenes. These feature signals are crucial for guiding attention in a bottom-up manner \cite{KatsukiConstantinidis2014}. This suggests that bottom-up attention may still play a role in the representation of global shape.

Research can manipulate the shape saliency of skeletal structures by proportionally altering the distance between the two vertical axes of the shape skeleton according to a gradient \cite{Ayzenberg Lourenco, 2019}. By systematically investigating dynamic changes in regional activation, researchers can explore whether bottom-up attention influences the representation of global shape. Furthermore, studies utilizing natural objects \cite{Ayzenberg, Simmons, 2023} have demonstrated selective activation for global shape. It is noteworthy that information processing is jointly driven by two distinct attentional processes.

Existing evidence suggests that sensitivity to global shape is not influenced by task difficulty or the allocation of attentional resources \cite{Arsenovic2022}. Even from the perspective of large-scale cortical organization, the allocation of attention does not alter the sensitivity of the dorsal and ventral pathways to global shape, which remains most pronounced in areas V3b and IPS0, respectively \cite{Goldstein-Marcusohn2024}. Given that the representation of global shape appears unaffected by top-down attention, future research could employ an orthogonal experimental design driven by both features and tasks \cite{Lawrence2019}. Such an approach would allow researchers to control the salience of global shape to guide bottom-up attention, while simultaneously manipulating target tasks to guide top-down attention. This would facilitate an investigation into how object familiarity modulates the mechanisms of global shape representation and integration within the dorsal pathway.

Despite substantial evidence supporting the dorsal pathway's ability to independently represent the global shape of objects, some studies have found that patients with dorsal cortical damage but intact ventral cortices rarely exhibit significant object recognition deficits \cite{Goodale1994, Goodale2023, Milner2006}. If the representation of global shape is indeed critical for object recognition, why does this apparent contradiction exist? This discrepancy raises fundamental questions regarding the functional necessity of dorsal shape representations in everyday visual perception.

One perspective suggests that the stimuli used in these studies may be relatively familiar to the patients, allowing them to complete recognition and categorization tasks by relying solely on local features without recruiting the holistic shape representation information of the dorsal pathway (Ayzenberg & Behrmann, 2023b).

In a study by Ayzenberg and Blauch (2023), deep neural networks (DNNs) were trained using the same shape stimuli previously presented to patients in the work of Goodale and Meena (1994). The results demonstrated that a shallow feedforward network (CORnet-Z) could easily classify these stimuli into curved or straight categories. Such networks do not represent the global shape of an object; instead, they are extremely sensitive to local features (Baker et al., 2018, 2023; Jarvers & Neumann, 2023). These findings suggest that these objects can be successfully categorized even in the absence of holistic shape representations.

Research has demonstrated that when presented with images of object parts containing a small number of diagnostic features, humans can identify objects with high accuracy. Simultaneously, the ventral cortex exhibits high levels of activation during these tasks (Holzinger, 2019; Ullman, 2016). While certain computational models serve as optimal simulations for object recognition within the human ventral pathway (Ayzenberg & Behrmann, 2022a), both these models and the human ventral pathway's representation of local features are shaped by experience (Ayzenberg & Behrmann, 2024; Doerig, 2023). Consequently, local features possess high information availability when identifying familiar objects.

However, the information availability of local features no longer provides an advantage when encountering new or unfamiliar objects. In such cases, the sample sizes and training iterations required for novel object recognition are significantly larger than those needed by humans, and the resulting training performance remains comparatively poor (Zador, 2019). Studies on infants have shown that even at six months of age, infants can complete object categorization based on the global shape of an object after being exposed to only a single sample—a phenomenon known as one-shot learning (Ayzenberg & Lourenco, 2022).

The computational model that best matches infant behavioral performance is the skeletal model, which describes global shape (Ayzenberg & Lourenco, 2019). This suggests that in the absence of extensive perceptual experience or significant linguistic involvement, infants primarily rely on global shape to develop their object recognition capabilities.

While the visual world of infants is sparse in terms of object variety, their object classification capabilities can match or even surpass many state-of-the-art machine learning models \cite{Ayzenberg2024}. Current models may lack critical structural features or appropriate learning biases—such as samples that maximally display global shape structure—thereby limiting their rapid development in object recognition. Research has shown that integrating skeletal models into convolutional neural networks (CNNs) not only improves their performance in visual perception tasks but also brings their classification performance closer to that of humans \cite{Rezanejad2019}. Therefore, future research must incorporate constraints on the spatial arrangement of local features into model architectures to enhance their biological plausibility or their alignment with the dorsal stream.

To improve recognition accuracy and training efficiency, it is also necessary to integrate findings from infant development, computational modeling, and brain injury research. This will allow for an exploration of how the mechanisms emerging during learning and training influence the processing of global shape and how they resist interference. A key mechanism of adaptive visual object processing is the visual system's ability to maintain representations of current task-oriented targets even in the presence of distractors, interfering stimuli, or non-task goals \cite{Xu2024}. Representations of both targets and distractors coexist in areas such as the early visual cortex \cite{Christophel2018, Olmos-Solis2021, Rademaker2019, Xu2018, Xu2024}.

How does the visual system effectively resist interference to achieve representation of the current target? One view suggests that the visual system prevents interference by suppressing activity unrelated to the current task goal \cite{Olmos-Solis2021}. However, research by \cite{Xu2024} found that the key to resisting interference lies not in suppressing distractors, but in whether the representations of targets and distractors can be separated to avoid mutual interference. This is evidenced by the fact that changes in interfering stimuli or non-task goals do not affect the accuracy of target classification. Researchers believe that the representations of different information processed in parallel utilize "orthogonalization" to separate target and distractor representations.

In cases where such orthogonalized representations are absent, target classification accuracy is significantly affected by changes in interfering stimuli. This may reflect an encoding stage that faithfully mirrors the external environment rather than responding flexibly to task requirements. In this stage, the representation of interfering stimuli is much stronger than that of the target; however, at the representation level, there is no difference, suggesting that the system also inhibits interference to some extent when selecting and extracting information \cite{Xu2024}. It is possible that the visual system simultaneously employs both inhibition and orthogonalization to resist interference, with the latter being more adaptive because it can accommodate representations of both target and distractor information. Notably, the experiments by \cite{Xu2024} only involved the continuous presentation of interfering images during a delay period and did not include the process of searching for and selecting target objects, which may explain why the study did not explicitly identify the presence of inhibition.

Future research could build upon these findings by increasing the number of search targets or the complexity of task processing. This would allow for a more detailed investigation into the dynamic mechanisms by which the brain employs orthogonalization and inhibition to counteract representational interference within the working memory system.

Furthermore, research indicates that multiple brain regions involved in representation appear to follow a representational gradient. Specifically, anterior regions provide more abstract, goal- or task-relevant information, while posterior regions and the early visual cortex encode stimulus features such as category information. These regions can encode both task relevance and object categories \cite{Olmos-Solis2021}. As a functional integration platform, the posterior parietal cortex (PPC) can integrate information regarding object perceptual categories from the ventral pathway with task-relevance information from the prefrontal cortex. This integration forms a unified, goal-directed representation priority map, which prioritizes the processing of target information relevant to the current task while ignoring irrelevant distractors \cite{Olmos-Solis2021, Shenhav2024}.

Therefore, future research needs to further investigate the influence of endogenous memory information from the ventral pathway on adaptive visual object representation. While Xu \cite{Xu2018a, Xu2018b} demonstrated how cognitive functions appropriately integrate into the adaptive visual representation of objects, this visual information primarily originates from the ventral pathway's processing of current external stimuli (exogenous information). In fact, the PPC can also retrieve and extract relevant long-term stored episodic and semantic memory information from the medial temporal lobe (MTL) \cite{Brown2018, Humphreys2022, Ramanan2018} and the anterior temporal lobe (ATL) of the ventral pathway \cite{Humphreys2022}. Specifically, the lateral intraparietal sulcus (LIPS) and the angular gyrus (AG) can temporarily buffer and integrate this information \cite{Humphreys2021, Kuhnke2023, Ramanan2018}. As part of the frontoparietal control network, these regions are regulated by top-down executive control from the prefrontal cortex (PFC), allowing for the selection and manipulation of endogenous information based on current task demands \cite{Humphreys2022, Humphreys2021, Sestieri2017}.

The processing of endogenous information is functionally similar to the processing of exogenous information. Given that the AG is anatomically adjacent to the IPS, these regions may represent different components of the adaptive visual object representation system. However, little is known about the mechanisms by which brain regions involved in processing endogenous information participate in the adaptive visual representation of exogenous information \cite{Xu2018a, Xu2018b}. Research suggests that the network centered on the AG for retrieving endogenous memory information and the network centered on the IPS for searching external perceptual input exhibit a functional competitive relationship \cite{Sestieri2017, Sheremata2018}. Similar competitive dynamics also exist between the anterior, posterior, and middle subregions of the IPS \cite{Humphreys2020, Humphreys2022}. Connectivity analyses further reveal that the posterior IPS is linked to extensive areas of the ventral pathway (including the primary visual cortex), while the middle IPS can receive both current perceptual information and long-term episodic and semantic memory information \cite{Humphreys2022}. Future research is necessary to clarify how the PPC integrates this multimodal spatiotemporal information input \cite{Humphreys2022}.

The relationship between operations is crucial for explaining the adaptive visual representation mechanisms for internally and externally generated information. Understanding the developmental mechanisms of the dorsal and ventral pathways, as well as the impact of their interaction, is fundamental to this field. Research on humans and non-human primates indicates that the developmental rates of the dorsal and ventral pathways are asynchronous \cite{Stiles2020}. For instance, studies on infants (under one year old) show that the dorsal pathway typically develops structurally and functionally earlier than the ventral pathway \cite{Ciesielski2019, 2021}. Conversely, in older children—such as those in preschool and school-age stages—the ventral pathway develops more rapidly than the dorsal pathway \cite{Vinci-Booher2022}.

This developmental pattern may be related to the fact that infants primarily utilize blurred vision and sensitivity to motion, prioritizing the development of global shape representation capabilities \cite{AyzenbergBehrmann2024}. Specifically, the dorsal magnocellular pathway develops earlier than the ventral parvocellular pathway \cite{Hammarrenger2003}, which may lead infants to categorize objects based on global shapes represented by the dorsal pathway rather than local features \cite{AyzenbergLourenco2022}. As visual experience with objects accumulates, the developmental speed of the ventral pathway accelerates, leading to the emergence of distinct category-selective regions for visual objects. This shift results in older children exhibiting a ventral pathway advantage in visual processing \cite{Srihasam2014}.

How, then, does the asynchrony in the development of the dorsal and ventral pathways affect their interaction in object representation? Research suggests that the processing advantage of the ventral pathway can drive the development of the dorsal pathway. For example, studies have found that when perceiving objects formed by ordered rotating light points, children aged 5–6 show increased activation in the dorsal pathway compared to adults (aged 20–30), while activation in the ventral lingual gyrus is also enhanced. This suggests that children’s processing of spatial motion information is still immature and requires support from object representations in the ventral pathway \cite{Klaver2008}. Furthermore, research has found that by age 8, the development of white matter fiber tracts within the ventral pathway has reached adult levels, whereas fiber tracts within the dorsal pathway remain immature and require a longer developmental trajectory \cite{Vinci-Booher2022}. Given that sensory input is a primary driver of brain white matter development \cite{MaurerLewis2018}, and that visual information is transmitted from the ventral pathway to the dorsal pathway via connecting fiber tracts \cite{Vinci-Booher2022}, it follows that the microstructural development of the dorsal pathway may require signals from the ventral pathway.

However, few studies have directly investigated whether the early development of the dorsal pathway during infancy drives the subsequent development of the ventral pathway. In human infants (0–6 months old), the dorsal pathway develops faster than the ventral pathway in terms of fiber tract myelination and synaptogenesis. This may be linked to the rapid development of motion perception capabilities in the infant dorsal pathway, which can support object recognition in the ventral pathway \cite{AyzenbergBehrmann2024, 2021}. As the frontoparietal attention network gradually develops within the first 6 months after birth, infants' fixation time on simple stimuli (such as black-and-white geometric patterns) decreases, while fixation time on complex stimuli and their features (such as faces and objects) increases, accompanied by higher activation levels in the ventral pathway \cite{Reynolds2015}. This suggests that the enhancement of ventral pathway function may be supported and driven by the dorsal attention network.

In summary, the dorsal and ventral pathways exhibit a dynamic, bidirectional interaction throughout individual development. This manifests as the early development of the dorsal pathway driving and supporting the later development of the ventral pathway during infancy, followed by the gradual maturation of the ventral pathway supporting the continued development of the dorsal pathway during childhood.

Research in this area must examine how the bidirectional interactions between the dorsal and ventral pathways during infancy and childhood influence the functional integration of these pathways in the processing of visual object invariance and adaptation.

参考文献

Chen, (2023). Bandwidth cortical representations multiple faces.

Cerebral Cortex, (18), Alizadeh, A.-M., Dromme, Verhoef, B.-E., Janssen, (2018).

Caudal intraparietal sulcus three-dimensional vision: combined functional magnetic resonance imaging single-cell study.

Neuroimage, Arsenovic, Ischebeck, Zaretskaya, (2022).

Dissociation between attention-dependent spatially specific illusory shape responses within topographic areas posterior parietal cortex.

Journal Neuroscience, (43), Ayzenberg, Behrmann, (2022a). brain's ventral visual pathway compute object shape?

Trends in Cognitive Sciences, 26 (12), 1119 – 1132.

Ayzenberg, Behrmann, (2022b). dorsal visual pathway represents object-centered spatial relations object recognition.

Journal Neuroscience, (23), Ayzenberg, Behrmann, (2023a). expanded neural framework shape perception.

Trends

in Cognitive Sciences, 27 (3), 212 – 213.

Ayzenberg, Behrmann, (2023b). where, what, object recognition.

Trends

Cognitive Sciences, 27 (4), 335-336.

Ayzenberg, Behrmann, (2024). Development visual object recognition.

Nature Reviews Psychology, Ayzenberg, Blauch, Behrmann, (2023).

Using neural networks address

object recognition. PsyArXiv

Ayzenberg, Chen, Yousif, Lourenco, (2019). Skeletal representations shape human vision:

Evidence pruned medial model. Journal Vision, Ayzenberg, Kamps, Dilks, Lourenco, (2022).

Skeletal representations shape human visual cortex.

Neuropsychologia, Ayzenberg, Lourenco, (2022). Perception object global shape described model skeletal structure human infants.

Elife, e74943. Ayzenberg, Lourenco, (2019). Skeletal descriptions shape provide unique perceptual

information for object recognition. Scientific Reports, 9 (1) , 9359 .

Ayzenberg, Simmons, Behrmann, (2023). Temporal asymmetries interactions between dorsal ventral visual pathways during object recognition.

Cerebral Cortex Communications, Baker, Garrigan, Phillips, Kellman, (2023).

Configural relations humans convolutional neural networks.

Frontiers Artificial Intelligence, Baker, Erlikhman, Kellman, (2018). convolutional networks classify based global object shape.

Computational Biology, e1006613 Berryhill, Olson, (2008). right parietal critical visual working memory.

Neuropsychologia, Bettencourt, (2016a). Decoding content visual short-term memory under

distraction occipital parietal areas. Nature neuroscience, Bettencourt, (2016b).

Understanding location- feature-based processing along human intraparietal sulcus.

Journal Neurophysiology, Borra, Luppino, (2017). Functional anatomy macaque temporo-parieto-frontal connectivity.

Cortex, Bracci, Daniels, Beeck, (2017). context overrules object-and category-related representational content human parietal cortex.

Cerebral ortex, Brown, Rissman, Chow, Uncapher, Wagner, (2018).

Differential medial temporal parietal cortical contributions real-world autobiographical episodic autobiographical semantic memory.

Scientific Reports, Budisavljevic, Dell'Acqua, Castiello, (2018).

Cross-talk connections underlying dorsal ventral stream integration during actions.

Cortex, Bullock, Takemura, Caiafa, Kitchell, McPherson, Caron, Pestilli, (2019).

Associative white matter connecting dorsal ventral posterior human cortex.

Brain Structure Function, Chen, Snow, Culham, Goodale, (2018).

Elongation Tool-Specific activation connectivity dorsal ventral visual streams?

Cerebral ortex, Christophel, Iamshchinina, Allefeld, Haynes, J.-D. (2018).

Cortical specialization attended versus unattended working memory.

Nature euroscience, Ciesielski, Stern, Diamond, Khan, Busa, Goldsmith, Rosen, (2019).

Maturational changes human dorsal ventral visual networks.

Cerebral ortex, (12), Collins, Freud, Kainerstorfer, Behrmann, (2019).

Temporal ynamics rocessing ifferentiate ontributions orsal entral isual athways.

Journal

C ognitive N euroscience, 31 (6), 821 – 836.

Conci, llner, Leszczynski, ller, (2011). time-course global local attentional guidance Kanizsa-figure detection.

Neuropsychologia, Destler, Singh, Feldman, (2019).

Shape discrimination along morph-spaces. Vision esearch, Doerig, Sommers, Seeliger, Richards, Ismael, Lindsay, Kietzmann, (2023). neuroconnectionist research programme.

Nature Reviews Neuroscience, Felleman, Burkhalter, Essen, (1997).

Cortical connections areas macaque monkey extrastriate visual cortex.

Journal Comparative Neurology, Fitzgerald, Freedman, Assa, (2011).

Generalized associative representations parietal cortex.

Nature euroscience, Freedman, Ibos, (2018). integrative framework sensory, motor, cognitive functions posterior parietal cortex.

Neuron, Freud, Behrmann, (2020). Altered large-scale organization shape processing visual agnosia.

Cortex, Freud, Behrmann, Snow, (2020). dorsal cortex contribute perception?

Mind, Freud, Culham, Plaut, Behrmann, (2017). large-scale organization shape

processing ventral dorsal pathways. Elife, e27576.

Freud, Ganel, Shelef, Hammer, Avidan, Behrmann, (2017).

Three-dimensional representations objects dorsal cortex dissociable those ventral cortex.

Cerebral ortex, Freud, Plaut, Behrmann, (2016). appening orsal isual athway.

Trends in Cognitive Sciences, 20 (10), 773 – 784.

Garcea, Chen, Vargas, Narayan, Mahon, (2018). Task- domain-specific modulation functional connectivity ventral dorsal object-processing pathways.

Brain Structure Function, Gillebert, Mantini, Thijs, Sunaert, Dupont, Vandenberghe, (2011).

Lesion evidence critical intraparietal sulcus spatial attention.

Brain, Goldstein-Marcusohn, Asaad, Asaad, Freud, (2024). large-scale organization shape processing ventral dorsal pathways dissociable attention.

Cerebral ortex, bhae221. Goodale, Jakobson, Milner, Perrett, Benson, Hietanen, (1994). nature limits orientation pattern processing supporting visuomotor control visual

form agnosic. Journal of C ognitive N euroscience, 6 (1), 46 – 56.

Goodale, Meenan, lthoff, Nicolle, Murphy, Racicot, (1994).

Separate neural pathways visual

analysis

object shape perception prehension. Current Biology, Goodale, Milner, (2023).

Shape perception require dorsal stream processing.

Trends in Cognitive Sciences, 27 (4), 333 – 334.

Grill-Spector, Kushnir, Edelman, Avidan, Itzchak, Malach, (1999).

Differential processing objects under various viewing conditions human lateral occipital complex.

Neuron, Goller, Ansorge, (2021). Simple shapes guide visual attention based their global outline global orientation contingent search goals.

Journal Experimental Psychology: Human Perception Performance, (11), Leclerc, Dapello, Madry, Dicarlo, (2022).

Adversarially trained neural representations already robust biological neural representations.

International

Conference on Machine Learning, 162 , 8072 – 8081.

Hammarrenger, Lepor Labrosse, Guillemot, J.-P., M.-S. (2003).

Magnocellular parvocellular developmental course infants during first life.

Documenta Ophthalmologica, Holzinger, Ullman, Harari, Behrmann, Avidan, (2019).

Minimal recognizable configurations elicit category-selective responses higher order visual cortex.

Journal ognitive euroscience, Humphreys, Jackson, Lambon Ralph, (2020).

Overarching principles dimensions functional organization inferior parietal cortex.

Cerebral ortex, (11), Humphreys, Jung, Lambon Ralph, (2022). convergence divergence episodic semantic functions across lateral parietal cortex.

Cerebral ortex, (24), Humphreys, Ralph, Simons, (2021). unifying account angular gyrus contributions episodic semantic cognition.

Trends Neurosciences,

Ibos, Freedman, (2016). Interaction between spatial feature attention posterior parietal cortex.

Neuron, Jagadeesh, Gardner, (2022). Texture-like representation objects human visual cortex.

Proceedings National Academy Sciences, (17), e2115302119.

Janssen, Srivastava, Ombelet, Orban, (2008). Coding shape position macaque lateral intraparietal area.

Journal Neuroscience, (26), Jarvers, Neumann, (2023).

Shape-selective processing networks: ntegrating evidence perceptual integration.

Frontiers Computer Science, Jitsuishi, Yamaguchi, (2020).

Identification distinct association fiber tract IPS-FG connect intraparietal sulcus areas fusiform gyrus white matter dissection tractography.

Scientific Reports, Kastner, Chen, Jeong, Mruczek, (2017). brief comparative review primate posterior parietal cortex: novel hypothesis human toolmaker.

Neuropsychologia, Katsuki, Constantinidis, (2014).

Bottom-up top-down attention: ifferent processes overlapping neural systems.

Neuroscientist, Klaver, Lichtensteiger, Bucher, Dietrich, Loenneker, Martin, (2008).

Dorsal stream development motion structure-from-motion perception.

Neuroimage, Kleineberg, Dovern, Binder, Grefkes, Eickhoff, Fink, Weiss, (2018).

Action semantic knowledge Effective connectivity underlying neural networks.

Human Brain Mapping, Konen, Kastner, (2008). hierarchically organized neural systems object information human visual cortex.

Nature euroscience, Kravitz, Saleem, Baker, Ungerleider, Mishkin, (2013). ventral visual pathway: expanded neural framework processing object quality.

Trends Cognitive Sciences, Kuhnke, Chapman, Cheung, Turker, Graessner, Martin, Hartwigsen, (2023). angular gyrus semantic cognition: synthesis functional neuroimaging studies.

Brain Structure Function, Lawrence, Norris, Lange, (2019).

Dissociable laminar profiles concurrent bottom-up top-down modulation human visual cortex.

Elife, e44422. Lefco, Brissenden, Noyce, Tobyne, Somers, (2020).

Gradients functional organization posterior parietal cortex revealed visual attention, visual short-term memory, intrinsic functional connectivity.

Neuroimage, Lerner, Hendler, Malach, (2002). Object-completion effects human lateral occipital complex.

Cerebral ortex, Lescroart, Biederman, (2013). Cortical representation medial structure.

Cerebral ortex, Lima, Florentino, Fiorani, Soares, Schmidt, Neuenschwander, Gattass, (2023).

Cortical fundamental neural substrate visual representation.

Progress Neurobiology, Wang, Zhou, Ding, (2017). Perceptual integration rapidly activates dorsal visual pathway guide local processing early visual areas.

Biology, e2003646

Long, C.-P., Konkle, (2018). Mid-level visual features underlie high-level categorical organization ventral stream.

Proceedings National Academy Sciences, (38), E9015 E9024.

Long, Kuhl, (2018). Bottom-up top-down factors differentially influence stimulus representations across large-scale attentional networks.

Journal Neuroscience, (10), Mahon, (2023). Higher order visual object representations: functional

analysis

their perception action. Brown, Crosson, Haaland, (Eds.), handbook neuropsychology, Volume Neuroscience neuromethods. 138):

American Psychological Association. Maurer, Lewis, (2018).

Visual systems. (Eds.), neurobiology brain behavioral development 233).

London: Academic Press. Medina, Coslett, (2020). Impairments action perception after right intraparietal damage.

Cortex, Milner, (2017). visual streams interact other?

Experimental Brain Research, Milner, Goodale, (2006).

Disorders spatial perception visual control action.

Milner Goodale (Eds.), visual brain action(second edition) 120).

Oxford: Oxford University Press.

Natu, Rosenke, Querdasi, Kular, Lopez-Alvarez, Grill-Spector, (2021).

Infants cortex undergoes microstructural growth coupled myelination during

development. Communications Biology, 4 (1), 1191.

Nestmann, Karnath, H.-O., Rennig, (2022). ventral stream areas

viewpoint-invariant object recognition. Neuroimage, 251 , 119021.

Nestmann, Wiesen, Karnath, H.-O., Rennig, (2021). Temporo-parietal brain regions involved higher order object perception.

Neuroimage, Orhun, ones-Hinojosa, Middlebrooks, evik, Usseli, Bozkurt, (2024).

Topographic anatomy lateral surface parietal relationship white matter tracts.

Frontiers Neuroanatomy, Olmos-Solis, Loon, Olivers, (2021).

Content status: Frontal posterior cortical representations object category upcoming goals working memory.

Cortex, Rademaker, Chunharas, Serences, (2019). Coexisting representations sensory mnemonic information human visual cortex.

Nature euroscience, Ramanan, Piguet, Irish, (2018).

Rethinking angular gyrus remembering imagining future: contextual integration model.

Neuroscientist, Regev, Winawer, Gerber, Knight, Deouell, (2018).

Human posterior parietal cortex responds visual stimuli early peristriate occipital cortex.

European Journal Neuroscience, (12), Reynolds, (2015).

Infant visual attention object recognition. Behavioural esearch, Rezanejad, Downs, Wilder, Walther, Jepson, Dickinson, Siddiqi, (2019 Scene categorization contours:

Medial based salience measures [Paper presentation] IEEE/CVF Conference Computer Vision Pattern Recognition Beach,

California, United States .

Riddle, Hwang, Cellier, Dhanani, D'Esposito, (2019).

Causal evidence neuronal oscillations bottom attention.

Journal ognitive euroscience, Riddoch, Humphreys, Akhtar, Allen, Bracewell, Schofield, (2008). agnosias: istinctions between integrative agnosia.

Cognitive Neuropsychology, Ritz, Shenhav, (2024). Orthogonal neural encoding targets distractors supports

multivariate cognitive control. Nature Human Behaviour, 8 (5), 945 – 961.

Roumazeilles, Eichert, Bryant, Folloni, Sallet, Vijayakumar, Reveley, (2020).

Longitudinal connections organization temporal cortex macaques, great apes, humans.

Biology, e3000810. Schneider, Beste, Wascher, (2012). course bottom processes selective visual attention: study.

Psychophysiology, (11), Seideman, Stanford, Salinas, (2022). conflict between spatial selection evidence

accumulation in area LIP. Nature C ommunications, 13 (1), 4463.

Sereno, Maunsell, (1998). Shape selectivity primate lateral intraparietal cortex.

Nature, (6701), Sestieri, Shulman, Corbetta, (2017). contribution human posterior parietal cortex episodic memory.

Nature Reviews Neuroscience, Sheremata, Somers, Shomstein, (2018).

Visual short-term memory activity parietal reflects cognitive processes beyond attentional selection.

Journal Neuroscience, Srihasam, Vincent, Livingstone, (2014).

Novel domain formation reveals proto-architecture inferotemporal cortex.

Nature euroscience, (12), Stiles, Akshoomoff, Haist, (2020). development visuospatial processing.

Rubenstein Rakic (Eds.), Neural circuit cognitive development:

Comprehensive Developmental Neuroscience second edition 393).

Cambridge, Academic Press. Takemura, Kaneko, Sherwood, Johnson, Axer, Hecht, Leopold, (2024). prominent vertical occipital white matter fasciculus unique primate brains.

Current Biology, (16), Takemura, Pestilli, Weiner, (2019).

Comparative neuroanatomy: Integrating classic modern

methods

understand association fibers connecting dorsal ventral visual cortex.

Neuroscience Research, Taylor, (2022). Identifying neural mediating conscious object orientation

perception using fMRI MVPA. Cognitive Neuropsychology, 39 (1-2), 64 – 67.

Taylor, (2024). Using examine nonlinear mixed selectivity tuning category human brain.

Imaging Neuroscience, Theys, Romero, Loon, Janssen, (2015).

Shape representations primate dorsal visual stream.

Frontiers Computational Neuroscience, Theys, Srivastava, Loon, Goffin, Janssen, (2012).

Selectivity three-dimensional contours surfaces anterior intraparietal area.

Journal Neurophysiology, Thomas, Kveraga, Huberle, Karnath, H.-O., (2012).

Enabling global processing

simultanagnosia psychophysical biasing visual pathways.

Brain, Tseng, T.-Y., Muggleton, Tzeng, Hung, Juan, C.-H. (2010).

Posterior parietal cortex mediates encoding maintenance processes change blindness.

Neuropsychologia, Ullman, Assif, Fetaya, Harari, (2016).

Atoms recognition human computer vision. Proceedings National Academy Sciences, (10), Loon, Olmos-Solis, Fahrenfort, Olivers, (2018).

Current future goals represented opposite patterns object-selective cortex.

Elife, e38677. Vaziri-Pashkam, Taylor, (2019). Spatial frequency tolerant visual object representations human ventral dorsal visual processing pathways.

Journal ognitive euroscience, Vaziri-Pashkam, (2017).

Goal-directed visual processing differentially impacts human ventral dorsal visual representations.

Journal Neuroscience, (36), Vaziri-Pashkam, (2019). information-driven 2-pathway characterization occipitotemporal posterior parietal visual object representations.

Cerebral ortex, Vinci-Booher, Caron, Bullock, James, Pestilli, (2022).

Development white matter tracts between within dorsal ventral streams.

Brain Structure Function, Wang, Janini, Konkle, (2022).

Mid-level feature differences support early animacy object distinctions:

Evidence electroencephalography decoding. Journal ognitive euroscience, Wang, Zhou, Chen, Huang, (2023). subcortical magnocellular pathway responsible processing topological properties objects: transcranial magnetic stimulation study.

Human Brain Mapping, Wilder, Rezanejad, Dickinson, Siddiqi, Jepson, Walther, (2019).

Local

contour symmetry facilitates scene categorization.

Cognition, 182 , 307 – 317.

Wurm, Caramazza, (2022). pathways action object recognition.

Trends

in Cognitive Sciences, 26 (2), 103 – 116.

(2010). neural task-irrelevant features object-based processing.

Journal Neuroscience, (42), (2018a). posterior parietal cortex adaptive visual processing.

Trends Neurosciences, (11), (2018b). visual systems:

Invariant adaptive visual information representations

in the primate brain. Annual Review of Vision Science, 4 (1), 311 – 336.

(2020). Revisit sensory storage account visual working memory.

Visual

Cognition, 28 (5-8), 433 – 446.

(2023a). Global object shape representations primate brain.

Trends Cognitive Sciences, (2023b). Parietal-driven visual working memory representation occipito-temporal cortex.

Current Biology, (20), (2024). human posterior parietal cortices orthogonalize representation different streams information concurrently coded visual working memory.

Biology, (11), e3002915.

Jeong, (2015). contribution human superior intraparietal sulcus visual short-term memory perception.

Jolicoeur, Lefebvre, Martinez-Trujillo (Eds.), Mechanisms sensory working memory: attention performance York:

Academic Press. Vaziri-Pashkam, (2022). Understanding transformation tolerant visual object representations human brain convolutional neural networks.

Neuroimage, Yang, (2020). Domain-specific functional coupling between dorsal ventral systems during action perception.

Scientific Reports, Shim, (2017). Occipital, parietal, frontal cortices selectively maintain task-relevant features multi-feature objects visual working memory.

Neuroimage, Zachariou, Nikas, Safiullah, Gotts, Ungerleider, (2017).

Spatial mechanisms within dorsal visual pathway contribute configural processing faces.

Cerebral ortex, Zador, (2019). critique learning artificial neural networks learn

animal brains. Nature C ommunications, 10 (1) , 3770 .

Dorsal-ventral pathway interactions visual object representation Houde Faculty Psychology, Shandong Normal University, Jinan China School

Psychology, Nanjing Normal University, Nanjing 210097, China)

Abstract

visual system processes object information through dorsal ventral pathways. ventral pathway primarily responsible "what" object recognition, whereas dorsal pathway mainly specializes "where", dealing visuospatial motion processing.

However, extensive reciprocal neural connections between these pathways suggest significant functional interactions visual object representation. hand, ventral pathway cannot independently compute object's global shape. integration global shape information, represented dorsal pathway, local feature information, processed ventral pathway, necessary support invariant object recognition. other hand, during goal-directed thought behavior, dorsal pathway dynamically extract maintain object information ventral pathway facilitate adaptive visual processing. former constitutes primarily feature-driven (bottom-up) functional integration, while latter represents predominantly task-driven (top-down) functional integration.

Future research should further explore several issues: influence attention dorsal pathway's representation global shape; mechanisms which object familiarity modulates integration global shape local features; visual working memory resists interference maintain

processing goal-relevant stimuli; impact endogenous memory adaptive object representation; developmental trajectories dorsal ventral pathways their implications their functional interactions.

Keywords

isual object representation orsal-ventral pathway interaction, nvariant visual object processing, daptive visual object processing