[Verse 1] In twenty-sixteen a paper dropped knowledge so profound Lillicrap showed us feedback weights don't need to be crowned Random connections flowing backward through the neural maze Still teach the network proper moves in computational ways The brain don't need symmetric paths like backprop demands Nature builds with messy wires but learning still expands [Chorus] Random feedback, random feedback, breaks the symmetry chains Direct alignment, direct alignment, rewires learning's reins No perfect mirrors in biology, yet gradients still flow These cortical columns dance and grow, that's how the patterns go [Verse 2] Nøkland stepped up that same year with alignment so direct Skip the backward pass entirely, connect output to reflect Each hidden layer gets its signal straight from error's call No need to propagate precisely, feedback weights enthrall The mathematics still converge though pathways seem bizarre Biological plausibility brings learning from afar [Chorus] Random feedback, random feedback, breaks the symmetry chains Direct alignment, direct alignment, rewires learning's reins No perfect mirrors in biology, yet gradients still flow These cortical columns dance and grow, that's how the patterns go [Bridge] Bartunov came in twenty-eighteen to scale these ideas wide Testing algorithms against the tide of deeper architectures Can random weights survive the test when networks multiply The scalability assessment shows where bio-learning flies [Verse 3] Three papers paint a picture of how cortex might compute Without the weight transport problem that makes backprop mute Synaptic feedback randomly wired but functionally sound Local learning rules distributed across neural ground The columns process information in parallel streams While feedback shapes the learning through biological schemes [Verse 4] Evolution carved these circuits through millions of years past No central planner needed when learning rules amass Hebbian updates mixing with the random weight cascade Neurons wire together when their patterns masquerade The beauty of these systems lies in their robust design Approximating gradients through feedback so divine [Chorus] Random feedback, random feedback, breaks the symmetry chains Direct alignment, direct alignment, rewires learning's reins No perfect mirrors in biology, yet gradients still flow These cortical columns dance and grow, that's how the patterns go [Outro] From Lillicrap to Nøkland then Bartunov's testing ground The mathematics of the cortex in these papers can be found Random weights supporting learning, alignment showing way Biological computation's here to stay
← Neural Phantom's Random Dance | Misaligned Pathways Still Can Teach →