Neuroscience paper shows shunting inhibition boosts local credit assignment
Biological dendrites can approximate backprop with shunting inhibition, but still lag 5-6%
A new paper by Houman Safaai, Maceo Richards, and Bernardo L. Sabatini explores how biological neurons solve the credit assignment problem using shunting inhibition and dendritic branching. The team built conductance-based networks with E/I synapse banks and tree-structured branch-to-soma coupling. They show that exact gradients decompose into local eligibility terms (presynaptic activity, driving force, input resistance) and a fast non-local term that transports somatic error through dendritic gains. This turns local learning into a compression problem: shaping the compartment-error field to match global feedback.
Testing their hypothesis, the researchers found that shunting inhibition helps reshape the error field to better match low-rank or path-structured feedback. However, under 5-factor (5F) feedback with nonnegative conductances, their 'shunting LocalCA' algorithm remained 5–6 percentage points below matched backpropagation on MNIST, Fashion-MNIST, and figure-ground MNIST datasets. The work highlights how E/I conductance and dendritic geometry can influence credit-signal geometry, but also reveals that feedback-field fidelity remains a major bottleneck for biologically plausible learning algorithms.
- Factorization: gradients split into local eligibility (presynaptic activity, driving force, input resistance) and path-specific error via dendritic gains.
- Shunting inhibition reshapes compartment-error field to better approximate global feedback signals like low-rank or path-structured targets.
- Despite improvements, shunting LocalCA lags 5–6 percentage points behind backprop on MNIST variants, showing feedback fidelity is still a bottleneck.
Why It Matters
Bridges neuroscience and AI by showing how dendrites can implement local learning, but reveals fundamental limits of biologically plausible algorithms.