Korva is a scroll of the internet’s best ideas.
Bear, tiger, truck: one fear
From pixels to abstractions
At the retina, a charging bear is just a bunch of pixels, a blobby black shape growing as it comes closer. The brain builds more abstract representations out of that, the way deep convolutional networks turn pixels into objects. A familiar face gets recognized on the left or right, near or far. Emotions work the same way: a bear, a tiger, or a big pickup truck coming at someone all funnel into a central abstract representation.
Fan-in, fan-out
Neuroscientists David Anderson and Ralph Adolphs call this a fan-in, fan-out architecture. The inputs are endless: seeing a bear, smelling one, hearing a twig snap, hearing someone yell that there is a bear. The outputs are just as varied: hide, run away, or defend yourself, depending on context.
Why not one reflex each
A reflex has a single input and a single behavior. Reflexes alone would need a separate one for the bear, the tiger, and every other large threat out there, which would not be efficient.
How do deep networks borrow from visual cortex?
David Hubel and Torsten Wiesel found in the 1950s and 1960s that cells in cat visual cortex respond to edges at particular orientations, with simple cells feeding into complex cells that tolerate shifts in position. They shared the 1981 Nobel Prize in Physiology or Medicine for it. That layered design inspired Kunihiko Fukushima's neocognitron in 1980. Yann LeCun's convolutional networks in the late 1980s refined the same idea.
How does the brain recognize a face from any angle?
In macaque monkeys, a chain of face-responsive patches runs through the temporal lobe. A 2010 study in Science by Winrich Freiwald and Doris Tsao found that the earlier patches respond to faces at particular head angles. A patch further along responds to the same individual across different views. The step from view-specific to view-independent coding is a stage-by-stage build of an abstract representation from raw visual input.
More from Huberman Lab
