Where artificial and biological vision diverge (and why it matters)

acid house boom bap, dembow balkan brass band, afro-jazz carnatic · 3:58

Listen on 93

Lyrics

[Verse 1]
The camera clicks and captures light, but that's where likeness ends
CNN flows straight ahead while brain circuits bend
Feedforward stacks march single file through layers clean and neat
But biological vision loops back, makes the cycle complete

[Chorus]
Where silicon meets flesh, the pathways split in two
Feedforward versus recurrent loops, spikes versus values smooth
Backprop trains the artificial mind with gradients flowing back
While Hebbian rules and local codes keep biology on track

[Verse 2]
In primate IT cortex, some images take their time
The settling loop keeps working when first pass can't define
Those tricky visual puzzles that make the monkeys pause
Are exactly where the CNNs break their feedforward laws

[Chorus]
Where silicon meets flesh, the pathways split in two
Feedforward versus recurrent loops, spikes versus values smooth
Backprop trains the artificial mind with gradients flowing back
While Hebbian rules and local codes keep biology on track

[Verse 3]
Artificial neurons whisper numbers soft and continuous
But real neurons fire in bursts, timing synchronous
Rate codes tell us how much, but spikes encode the when
Latency and synchrony dance in the temporal den

[Bridge]
Add some recurrence to the net, CORnet-R shows the way
Closing gaps between machine and brain processing day by day
But adversarial attacks still fool the pixel-reading eye
While human vision sees the shape that textures can't deny

[Verse 4]
The global gradient flows backward through each weighted connection
But neurons can't send error signals in reverse direction
STDP and BCM rules, three factors gate the learning
Local plasticity solves the puzzle that researchers are discerning

[Verse 5]
Object recognition scales through the ventral visual stream
V1 detects the edges, IT holds the semantic dream
But cortical columns organize in ways we've yet to code
While artificial layers stack like freight along a road

[Chorus]
Where silicon meets flesh, the pathways split in two
Feedforward versus recurrent loops, spikes versus values smooth
Backprop trains the artificial mind with gradients flowing back
While Hebbian rules and local codes keep biology on track

[Outro]
Fragile networks stumble on perturbations microscopic
While robust biological vision stays philosophic
The divergence shows us clearly what the feedforward stack is missing
Recurrent loops and spiking codes are what we should be wishing

[Chorus]
Where silicon meets flesh, the pathways split in two
Feedforward versus recurrent loops, spikes versus values smooth
Backprop trains the artificial mind with gradients flowing back
While Hebbian rules and local codes keep biology on track

← The motifs artificial networks took from biology | The benchmark that scores the correspondence →