The motifs artificial networks took from biology

hip hop dubstep, doo-wop classical · 4:22

Listen on 93

Lyrics

[Verse 1]
Back in nineteen-eighty Fukushima saw the light
Hubel and Wiesel's simple cells caught his sight
He built the Neocognitron, layer upon layer
Convolution meets pooling, nature's neural player
Filter then pool, that's the hierarchy's call
From retina's blueprint to the artificial hall

[Chorus]
Biology's whispers in silicon dreams
Convolution and pooling, simple and complex schemes
From edges to textures, parts to the whole
Gabor detectors emerge from the code
The brain's ancient wisdom in networks reborn
Teaching machines how vision is worn

[Verse 2]
First layer convergence, a remarkable sight
Gabor-like filters learning edge detection's might
Oriented lines appear without human design
The same family V1 uses, crossing that line
Natural images train what evolution knew
Artificial features matching biology's view

[Chorus]
Biology's whispers in silicon dreams
Convolution and pooling, simple and complex schemes
From edges to textures, parts to the whole
Gabor detectors emerge from the code
The brain's ancient wisdom in networks reborn
Teaching machines how vision is worn

[Verse 3]
Hierarchy climbs from V1 to IT's peak
V2, V4 stages that neural networks seek
Edges become textures, textures turn to parts
Parts build up objects, artificial arts
The correspondence loose but measurable and real
Abstraction's ladder that both systems feel

[Bridge]
Divisive normalization found its digital twin
Batch norm and layer norm, letting gradients win
Attention mechanisms weight by relevance computed
Gain modulation's concept in transformers rooted
Center-surround whitening from retina's DoG
Input decorrelation, breaking visual fog

[Verse 4]
Difference of Gaussians cleans the visual stream
Whitening preprocessing, the retina's scheme
Surround suppression keeps the signal pure
Contrast normalization, cortex signature
Every innovation has its biological thread
Ancient solutions in silicon spread

[Verse 5]
Sparse coding principles from cortex inspire
Neural efficiency sets networks on fire
Skip connections echo brain's parallel streams
Residual pathways fulfilling visual dreams
Feedback loops and recurrence now finding their place
Biology's recursion in silicon's embrace

[Chorus]
Biology's whispers in silicon dreams
Convolution and pooling, simple and complex schemes
From edges to textures, parts to the whole
Gabor detectors emerge from the code
The brain's ancient wisdom in networks reborn
Teaching machines how vision is worn

[Outro]
From Fukushima's vision to transformers today
Biology guides us along neural pathways
Nature's algorithms refined through time
Inspire the networks in silicon's chime

← Cortical magnification | Where artificial and biological vision diverge (and why it matters) →