Hypercolumn: Ocular Dominance & Orientation Columns

The primary visual cortex (V1) alone is insufficient for object recognition. Additional specialized areas (V2, V3, V4, V5) contain their own visual field maps and communicate extensively. V1 sorts information into forms, colors, movements, and spatial orientation, sending signals to specialized areas. V2-V3 process forms, V4 processes colors, and V5 processes movement direction. After initial processing, information splits into two pathways: the ventral pathway (what pathway) for object recognition and the dorsal pathway (where pathway) for spatial location.

The visual cortex hierarchy beyond V1 includes V2, V3, V4, and MT (V5). V2 surrounds V1 and responds to illusory contours. V3 and V4 continue the ventral and dorsal streams. MT is located anterior to V4 and contains neurons responding to visual motion. The visual field representation follows retinotopic organization, with the fovea represented centrally. Modern neuroscience uses computational techniques to flatten the cortical surface and visualize retinotopic organization, revealing how different visual areas represent different aspects of visual information.

Visual processing extends beyond V1 to extra-striate areas. The dorsal pathway (parietal lobe) handles motion and spatial location; the ventral pathway (temporal lobe) handles object recognition. V2 handles segmentation and figure-ground. V3 processes moving objects. V4 processes color and form. V5 processes movement direction. Damage to V3-V5 impairs general recognition. Akinetopsia prevents perceiving movement, causing objects to appear to teleport. Red phenomenon perceives only movement. Antagonistic control manages selective attention, rejecting irrelevant stimuli to focus on relevant ones.

Visual information from V1 is processed through two main streams: the dorsal stream (where pathway) and ventral stream (what pathway). The dorsal stream originates from magnocellular layers of V1, projects through V2-V3 to V5, then to the dorsomedial temporal area (V5) and posterior parietal cortex, particularly the lateral intraparietal area for spatial attention. The ventral stream originates from parvocellular layers of V1, projects through V2 to inferior temporal areas, including V4 for color constancy and the fusiform face area for face recognition. These streams have multiple interconnections at various cortical levels.

Visual information flows from V1 (primary visual cortex) to secondary visual cortex (V2-V5, extra-striate cortex). Two streams emerge from V1: the dorsal stream projects to the parietal lobe as the 'how' pathway, guiding action and attention toward objects. The ventral stream projects to the temporal lobe as the 'what' pathway, identifying objects and connecting to memory systems for naming and recognition. This organization allows the brain to process both spatial interactions and object identification simultaneously.

Binocular disparity is the fundamental mechanism of binocular depth perception - the difference in images projected onto each retina when viewing a single object. The horopter is an imaginary sphere containing all points projecting to corresponding locations on both retinas with zero disparity. Objects inside the horopter (closer than focus point) and outside create disparity. Non-corresponding points on each retina signal depth - the greater the separation from the fovea, the farther apart objects are. Stereopsis is the perception of depth resulting from binocular disparity, allowing us to perceive three dimensions from two slightly different two-dimensional images presented to each eye.

Depth perception reconstructs the third dimension from two-dimensional retinal images through evolved mechanisms. Two categories exist: ocular motor cues (accommodation and vergence) and visual cues (binocular stereopsis and monocular cues including motion parallax, shading, interposition, size, and perspective). The stereoscope demonstrates stereopsis by presenting slightly different images to each eye, creating compelling 3D percepts. Bela Julesz's random dot stereograms prove stereopsis operates independently of recognizable features, relying solely on disparity cues. Approximately 5-10% of people lack stereopsis due to amblyopia or strabismus, with treatment effectiveness dependent on early intervention before age 8-10.

Binocular disparity is the difference between images on the two retinas, forming the basis of stereoscopic vision. The horopter is the surface where objects fall on corresponding retinal points. Crossed disparity (objects closer than fixation) and uncrossed disparity (objects farther) indicate relative position. Julesz demonstrated that random dot stereograms can create depth perception using only binocular disparity, proving stereopsis is a fundamental mechanism. The correspondence problem involves matching features between retinal images to calculate disparity accurately.

Stereopsis is the perception of depth and three-dimensional structure obtained from visual information derived from two eyes, resulting from horizontal disparities between the images projected to each retina. This binocular vision process involves two distinct mechanisms: coarse stereopsis (qualitative stereopsis) for peripheral motion and spatial orientation, and fine stereopsis (quantitative stereopsis) for precise depth perception in the central visual field. The neural basis involves disparity-selective neurons in the visual cortex, with coarse stereopsis associated with the magnocellular pathway and fine stereopsis with the parvocellular pathway. Stereopsis is essential for tasks requiring precise depth judgment, such as threading needles, catching balls, and operating stereoscopic instruments, and is present in approximately 97% of the population.

Stereopsis is the perception of relative depth of objects based on binocular disparity. It is the ability to fuse images simultaneously when horizontal disparity falls within Panum's fusional area, resulting in appropriate depth perception. Stereopsis is the most advanced and important grade of binocular vision, representing the brain's ability to interpret slight differences between retinal images to create three-dimensional perception. The three grades of binocular vision are: simultaneous macular perception (seeing two dissimilar objects simultaneously), fusion (combining slightly dissimilar images), and stereopsis (perceiving depth). Each grade builds upon the previous one, with stereopsis being the final and most advanced stage.

HyperNetworks is a neural network architecture introduced at ICLR 2017 where a secondary network generates weights for a primary network, inspired by the biological concept of genes producing phenotypes. The HyperNetwork takes layer indices as input and generates corresponding weight matrices for each layer of the main network. When applied to Convolutional Neural Networks, HyperNetworks generate weights for each layer by taking layer indices as input, producing weight matrices that are applied to CNN layers through tensor operations. This approach enables weight sharing across layers while allowing different weights for different layers. The architecture can be static (fixed weights at initialization) or dynamic (different weights at different time steps). To address parameter explosion, a scaling factor is introduced to control weight magnitudes, enabling more efficient and adaptive neural network architectures.

Convolutional Neural Networks (CNNs) are designed for grid-like data such as images, inspired by the visual cortex organization. They consist of convolutional layers (using filters/kernels to extract local features), pooling layers (subsampling to reduce dimensionality), fully connected layers, and normalization layers. The convolution operation extracts visual features from 2D image regions, followed by pooling to retain important information while reducing computational load.

Convolutional Neural Networks (CNNs) incorporate strong priors about image structure that make them well-suited for computer vision: stationarity of statistics (image properties remain consistent across locations) and locality of pixel dependencies (nearby pixels are more related than distant ones). These assumptions reduce the number of parameters needed while maintaining effective learning capacity. Despite being relatively unknown for image recognition at the time, CNNs were argued to be ideal because their capacity could be controlled by varying depth and breadth, and their architectural assumptions aligned with known properties of natural images. This combination of flexibility and biological plausibility made CNNs the right choice for tackling large-scale image recognition problems.

Convolutional Neural Networks (CNNs) have revolutionized machine vision, making Deep Learning one of AI's hottest topics. Pioneered by Yann Lecun of NYU and Facebook's AI director, CNNs became the standard for vision projects. In early 2015, machines first beat humans at object recognition challenges. The ImageNet competition, starting in 2012 when Geoff Hinton's team won using a CNN, has had every winner since employ this architecture, dramatically reducing error rates. The convolutional layer uses a flashlight analogy: multiple flashlights (neurons) shine on an image wall, each searching for the same pattern in different regions. Combined, they form filters that detect patterns like edges or color contrasts. Multiple filters operate in parallel, creating 3D grids (e.g., 8x6x4) that scan the entire image efficiently.

Convolutional Neural Networks (CNNs), invented by Professor LeCun in 1988, are a neural network architecture inspired by the visual cortex of the brain. Designed specifically for handling natural signals like images, video, audio, and speech, CNNs have become foundational technology in modern AI. Applications include character recognition, handwriting recognition, reading zip codes, speech recognition, plant and insect identification, and bird song classification. They are widely used in consumer devices including smartphones and automotive safety systems.