Dendritic In-Context Learning in a Single-Layer Spiking Neural Network

What the math gives to ML

The paper identifies a concrete architectural mechanism for in-context learning in spiking networks: a persistent, multidimensional dendritic state whose recurrence performs leaky online Widrow–Hoff least-mean-squares updates. The transferable asset is not spiking itself, but embedding an explicit online optimizer in subthreshold compartment dynamics, eliminating the need for inference-time synaptic plasticity, attention, or deep stacks. This suggests a compact recurrent or neuromorphic module that stores a task-specific linear predictor in dendritic voltage and updates it from each labeled context example. The most direct test is to replace the hidden state of a small SNN or recurrent token processor with this dendritic LMS state and measure Garg-style ICL accuracy, stability across seeds, and energy or latency at fixed parameter count.