← Reference library

REFERENCE LIBRARY

Visual PathwayVisual Pathway

The full route light takes from the retina to the cortex, then splits into the "what / where" streams — plus one counterintuitive fact: the brain sends more signal "down" this path than it sends "up."

One trunk: retina → LGN → V1 → two streams

Vision isn't a single step of "the eye sees → the brain knows." The signal passes several stations, and every one re-encodes rather than simply forwarding:

Retina

Turns light into neural signal and already starts "computing" — not uniform exposure, but emphasizing contrast and edges (see below).

LGN (lateral geniculate nucleus)

A relay station in the thalamus. But it's far more than a "forwarder": about 90% of its input actually comes from cortical feedback, only about 10% from the eyes — gating and prediction begin right here.

V1 (primary visual cortex)

The first cortical visual area, home to orientation-selective cells (Hubel & Wiesel): different neurons are tuned to edges at different angles, like a layer of edge detectors.

Two streams

The ventral stream (→ temporal lobe) handles "what," the dorsal stream (→ parietal lobe) handles "where / how to use." Damage can dissociate them: some people can recognize a cup but can't grasp it accurately, others can grasp it but can't recognize it.

Receptive fields and retinotopy

Two organizing principles that run through the whole pathway:

Receptive field — each visual neuron only "watches" a small patch of the visual field. Low levels have a center-surround structure (excited when the center is bright and the surround dark), which is naturally an edge/contrast detector; the higher you go, the larger and more abstract the receptive field becomes (from edges → shapes → objects → faces). This hierarchy of "small receptive fields stacking into large ones, simple features combining into complex ones" is exactly the biological prototype of convolutional networks.

Retinotopy — adjacent positions in the visual field are adjacent on the cortex too. The brain literally lays "space" out as a map; many sensory systems use this kind of "topology-preserving" mapping (hearing is tonotopic, somatosensation is body-topographic).

Counterintuitive: feedback > feedforward

The point most easily simplified away in textbooks, and the most important: in the visual pathway, the number of feedback (top-down) connections exceeds that of feedforward (bottom-up) connections. That 90% of LGN input comes from the cortex is the plainest evidence. It means the brain is mainly "sending predictions" toward the senses rather than passively receiving — this anatomical fact is the physical bedrock of the whole "perception as inference / predictive processing" framework.