The seven conditions for gradient equivalence

cloud rap, classical surf · 3:17

Listen on 93

Lyrics

[Verse 1]
Seven rules to make the gradients align
When cortical columns compute and define
First condition, settling dynamics must converge
Spectral radius less than one, let the fixed point emerge
Contractive flow brings the system to rest
Mathematical stability puts us to the test

[Chorus]
Seven conditions, gradient equivalence
Settling, energy, phases for reference
Feedback pathways, learning rules contrastive
Small perturbations, architecture massive
Seven conditions, make the training work
When biology and backprop don't shirk

[Verse 2]
Energy function guides the neural dance
Scalar potential gives gradient descent a chance
Symmetric connections make the math clean
Or random feedback with alignment unseen
Derivable dynamics from a single well
Energy landscape has stories to tell

[Chorus]
Seven conditions, gradient equivalence
Settling, energy, phases for reference
Feedback pathways, learning rules contrastive
Small perturbations, architecture massive
Seven conditions, make the training work
When biology and backprop don't shirk

[Verse 3]
Two phases dancing, free and target states
Or compartments where the error awaits
Temporal switching or spatial divide
Both representations living side by side
Difference between them encodes what we need
Error signal plants the learning seed

[Bridge]
Consistent feedback must reach every weight
Symmetric exact or random approximate
Sign-symmetric falls between the extremes
Feedback alignment fulfills the dreams

[Verse 4]
Local learning rules with contrastive form
Target minus free becomes the norm
Delta W proportional to the difference
Hebbian plasticity with added significance
Small perturbation limit makes it precise
Beta approaches zero, mathematics suffice

[Verse 5]
When cortical circuits learn from experience
Credit assignment through neural inference
Backprop's precision meets biology's way
Gradient flow without backward relay
Inference networks solve the learning game
Forward and feedback working just the same

[Chorus]
Seven conditions, gradient equivalence
Settling, energy, phases for reference
Feedback pathways, learning rules contrastive
Small perturbations, architecture massive
Seven conditions, make the training work
When biology and backprop don't shirk

[Outro]
Architecture matching, depth correspondence
Recurrent unrolling with feedforward concordance
Tied weights connect the forward flow
Seven conditions make gradients glow

← Connection to neuroscience | When conditions partially hold →