Visual System Organization

The visual system processes information hierarchically, beginning with the retina and ending in the primary visual cortex (V1). Retinal ganglion cells possess center‑surround receptive fields, so a small spot of light elicits a larger response than a large spot because the inhibitory surround suppresses activity for broader stimuli. This basic arrangement sets the stage for increasingly complex feature extraction in the cortex.

Receptive Field Mapping

In the lateral geniculate nucleus (LGN) and early V1, neurons retain the center‑surround organization before developing orientation selectivity. Orientation‑selective neurons in V1 can be approximated as linear filters; simple cells respond to specific orientations and positions, while complex cells respond to orientation regardless of the exact position within the receptive field. The point here is that the small circle gives a bigger response than the big circle, illustrating the influence of inhibitory surrounds.

Primary Visual Cortex (V1) Organization

V1 is arranged retinotopically, with cortical magnification allocating more cortical area to the central visual field. Columns run perpendicular to the cortical surface, and electrode penetrations reveal consistent orientation preferences within each column. Ocular dominance columns segregate input from the left and right eyes, and layer 4C preserves the magnocellular and parvocellular streams from the LGN. Endstopping, a reduction in response when a stimulus exceeds a certain length, reflects an inhibitory surround that sharpens edge detection.

Computational Models of Vision

Simple cells are modeled as Gabor functions—the product of a sinusoid and a Gaussian—capturing their orientation and spatial frequency tuning. Complex cells follow the energy model: responses of even and odd symmetric simple cells are squared and summed, eliminating phase sensitivity while preserving orientation selectivity. Modern convolutional neural networks (CNNs) trained on object‑recognition tasks naturally develop early‑layer filters that resemble these Gabor‑like simple cells, linking artificial vision to biological mechanisms.

Spatial Frequency Channels

Images can be decomposed into sine‑wave gratings of varying frequencies, orientations, and phases. The contrast sensitivity function (CSF) is an inverted‑U shape, showing peak sensitivity at medium spatial frequencies. Adaptation to a specific spatial frequency for about one minute creates a localized “bite” in the CSF, indicating that independent spatial frequency channels exist. These adaptation effects are orientation‑specific, suggesting a cortical rather than retinal origin. High and low spatial frequency channels must interact for object recognition, as demonstrated by the Abraham Lincoln pixelation illusion, where both coarse and fine details contribute to the perception of the portrait.

Mechanisms and Explanations

A population code resolves ambiguity in single‑neuron responses; the peak of the distribution across many neurons determines the perceived feature. The tilt aftereffect illustrates this: prolonged exposure to an oriented grating reduces the responsiveness of neurons tuned to that orientation, shifting the population peak and causing a perceived tilt in the opposite direction. The energy model’s squaring and summation of even and odd simple cell responses removes phase dependence while retaining orientation selectivity. Spatial frequency channels, each tuned to a specific band (typically 1.3 to 2 octaves wide), process images in parallel, and adaptation experiments reveal their separable nature.