Computational model comparison (explanatory adequacy)

acoustic classical, russian ambient techno, grime surf rock, grunge new jack swing · 3:43

Listen on 93

Lyrics

[Verse 1]
Scientists built machines that see like we do
Deep networks trained on objects, layer by layer grew
They tested against neurons in macaque brains
Found something shocking in the data chains
V4 and IT cortex responses matched the code
Better than any hand-built model ever showed

[Chorus]
Brain-Score is the benchmark, measuring how well
Your model's representations can predict and tell
What neurons fire when images appear
Sufficiency's the game, but causation's never clear
You can match the data, match behavior too
But that don't mean the brain works just like you

[Verse 2]
Yamins and DiCarlo cracked the mystery wide
Convolutional layers mirror what's inside
From V1 through V2, up to higher ground
Each network layer maps where brain responses are found
CORnet added feedback loops to match the time
When signals flow backward, keeping perfect rhyme

[Chorus]
Brain-Score is the benchmark, measuring how well
Your model's representations can predict and tell
What neurons fire when images appear
Sufficiency's the game, but causation's never clear
You can match the data, match behavior too
But that don't mean the brain works just like you

[Bridge]
Evidence in its own right, these models stake their claim
Computational theories playing the prediction game
What it proves is powerful - your theory could be right
What it cannot prove is mechanism's inner sight
Necessary but not sufficient for the mechanistic case
Multiple roads can lead to the same neural place

[Verse 3]
From pixels to perception, hierarchies unfold
Edge detectors, curves, and shapes, the story's being told
Object recognition emerges from the stack
But implementation details, well, that knowledge we still lack
The transformation's captured, the function's crystal clear
But how the wetware does it remains a frontier

[Verse 4]
ResNet, VGG, and AlexNet compete
Each architecture striving to make predictions complete
ImageNet pretrained then fitted to the brain
Silicon circuits learning biological refrain
Yet questions linger in the depths of neural space
Do brains compute like GPUs in this computational race?

[Final Chorus]
Brain-Score is the benchmark, measuring how well
Your model's representations can predict and tell
What neurons fire when images appear
Sufficiency's the game, but causation's never clear
You can match the data, match behavior too
But the brain's true secrets are still breaking through

← Two-photon and calcium imaging (correlation, many cells, spatial layout) | How the methods combine into proof →