Visual receptive field

The jumping spider solved the visual hierarchy problem through independent evolution: large forward-facing principal eyes scan for high-resolution detail while smaller secondary eyes provide wide-field motion detection. Motion triggers a redirect of the principal eyes to the point of interest — coarse detection feeding into directed fine analysis. David Hubel and Torsten Wiesel found the same architecture in a cat's visual cortex in 1959, built not by evolution but discovered there, neuron by neuron, in a darkened lab in Baltimore.

The neuron had been silent for hours. David Hubel and Torsten Wiesel had been projecting spots of light onto a screen in front of an anesthetized cat — dark spots on light backgrounds, light spots on dark — trying to locate the stimulus that would make a neuron in the visual cortex respond. Nothing. Then, repositioning a glass slide in the projector, the shadow of its edge swept across the screen at an angle. The neuron fired a volley of spikes.

It was not the light that mattered. It was the edge.

The visual cortex, it turned out, was not built to register illumination levels. It was built to detect features. This is modularity at the neural level: independently processing oriented edges, positions, and shapes in separate specialized populations, then integrating the outputs into coherent perception. Each neuron — Hubel and Wiesel called them "simple cells" — responded to a specific oriented edge (vertical, horizontal, or diagonal) at a specific location on the retina. Move the same edge one degree left or rotate it five degrees and the cell went quiet. The neuron was a highly specific detector, tuned to one element of visual geometry.

The discovery from their 1959 paper, published in Journal of Physiology, was layered. Above the simple cells, a second population — complex cells — fired to oriented edges regardless of exact position within the receptive field. They pooled input from many simple cells sharing the same orientation preference. The complex cell didn't care where exactly the edge appeared; it cared what angle. Above that, hypercomplex cells responded to edges of specific lengths or to corners. The visual cortex was organized as a processing hierarchy: local position-specific detectors feeding into position-invariant detectors feeding into shape detectors.

The architecture extended to development. In experiments that Hubel and Wiesel later recognized as their most consequential, they covered one eye of a kitten during the first weeks of life. When uncovered after the critical developmental window had closed, the eye appeared anatomically intact — cornea, lens, retina all normal — but the cat was functionally blind in that eye. The corresponding cortical neurons had been colonized by the other eye's inputs. Sensory experience during a precise developmental window determined which neural connections survived. The visual cortex was not hardwired from birth; it was experience-dependent, but only during a narrow critical period.

Hubel and Wiesel received the 1981 Nobel Prize in Physiology or Medicine for this body of work. Their 1959 and 1962 papers are among the most cited in neuroscience.

The computational cascade was immediate and definitive. Kunihiko Fukushima's neocognitron, published in 1980, explicitly implemented the simple/complex cell hierarchy in an artificial neural network: S-cells (simple) feeding into C-cells (complex), each layer more position-invariant than the last. Yann LeCun refined this into the convolutional neural network architecture in the late 1980s and 1990s. Convolutional neural networks learn hierarchically organized feature detectors — the first layers learn edges and orientations, the middle layers learn textures and shapes, the deeper layers learn object-level structure. This is Hubel and Wiesel's architecture, implemented in silicon.

Every modern computer vision system — facial recognition, medical imaging, autonomous vehicle perception — runs on this discovered architecture. Path dependence runs from Hubel and Wiesel's 1959 structure through every convolutional neural network: LeCun and Fukushima built on discovered biology, and every subsequent architecture inherits their hierarchy. Hubel and Wiesel did not design it. They found it, in a cat's visual cortex, in 1959.

The jumping spider arrived at the same solution through entirely separate evolution. Salticidae have two distinct visual subsystems: large forward-facing principal eyes with a narrow high-resolution field that they physically scan by moving the retina itself, and smaller secondary eyes providing wide-field motion detection. The secondary eyes detect motion and direct the principal eyes to the point of interest for high-resolution analysis. This is the same hierarchical architecture — coarse detection feeding into directed fine analysis — that Hubel and Wiesel described. Two evolutionary paths, 500 million years apart, selecting the same computational structure for processing visual information.