Modeling intermediate-level features in V1 and V2 using natural images

The National Eye Institute awarded the University of California, Berkeley $143,031 on December 4, 2025, under the Vision Research program (CFDA 93.867) to computationally model the role of spatial phase in the selectivity and invariance of responses in primary visual cortex (V1) and secondary visual cortex (V2) to intermediate-level features in natural images. The modeling-driven research will investigate how phase-related computations operate across V1 and V2 and generate testable predictions to guide experimental design for V1 and V2 physiology studies. The work addresses how the visual system forms invariant and selective response properties between V1 and inferotemporal cortex by leveraging statistical properties of natural images to build ecologically-driven predictive models of brain function. Performance of work occurs in California through May 31, 2028.

Abstract

The goal of this proposal is to computationally investigate the role of spatial phase in the selectivity and invariance of responses of primary (V1) and secondary visual cortex (V2) to intermediate-level features in natural images. This modeling-driven approach will further our understanding of phase-related computations in V1 and V2, as well as provide testable predictions to guide experimental design for studies of V1 and V2 physiology. Behaviorally, primates are robustly invariant to transformations of visual objects and scenes [63,64]. For example, given an object under different novel viewing angles, illumination conditions, and positions, humans are gener- ally able to easily recognize an instance of the object. Although the invariant and selective response properties of individual V1 neurons have been relatively well-studied, how these properties are formed between V1 and inferotemporal cortex (IT) is not understood. IT has been proposed that the visual system has been adapted to the statistics of the signals in its environment [33, 65]. Therefore, constructing models of the brain that leverage statistical properties of natural images will provide ecologically-driven predictions that can be tested in experiments. Spatial phase and phase structure are perceptually important signals in the visual perception of structure and form [7-12]. To investigate responses to intermediate-level properties in V1 and V2, such as elongated contours, corners, and boundaries, IT is essential to approach the problem using a normative model that leverages phase statistics of natural images. V1 complex cells are known to be relatively phase-invariant, as modeled in the standard energy model of COM- plex cells [35]. Although the phase-sensitive V1 simple cells have been found to project to V2 [16], how phase information is used in the ventral stream is not clear [17,18]. Constructing a model that explicitly factorizes phase and amplitude responses of units, with respect to natural image stimuli will provide hypotheses for how phase is represented in V1 and V2. This program will develop a first and second layer independently, each optimized to learn the phase and amplitude- based factors that best describe the stimuli. First, the preliminary first layer V1 model will be refined and tested for robustness, then compared to V1 data along with a baseline [22]. Then, after analysis of the joint phase statistics of the V1 unit responses, a second layer will be constructed from those dependencies. The unit invariances and selectivities in the model will provide testable experimental predictions. This modeling work will be performed using the computational resources in prof. Bruno olshausen's theoretical neuroscience group at the university of california, berkeley. IT will be supported by the theoretical neuroscience expertise of prof. Olshausen, and the physiological and computational expertise of collaborator prof. Timothy oleskiw at the university of regina. If successful, this project could lead to novel computational tools to investigate invariances and selectivity in V1 and V2, which will contribute to our understanding of how the primate brain forms the high-level percepts of shape and form.