| name | dendritic-in-context-learning-snn |
| description | DendriCL methodology for in-context learning in single-layer spiking neural networks using dendritic compartment dynamics. Use when: implementing ICL in biologically-plausible SNNs, designing compartmental spiking architectures, studying online LMS in dendrites, or building seed-stable ICL at super-dimensional task complexity. arXiv: 2607.02289 |
Dendritic In-Context Learning in Spiking Neural Networks (DendriCL)
Source
Paper: "Dendritic In-Context Learning in a Single-Layer Spiking Neural Network" (arXiv: 2607.02289, July 2026)
Core Discovery
In-context learning (ICL) does NOT require:
- Attention mechanisms
- Architectural depth
- Inference-time synaptic plasticity
A single dendritic compartment with online-LMS dynamics is sufficient.
Key Insight
Prior SNN designs fail the Garg-2022 ICL benchmark because they route adaptation through inference-time synaptic plasticity, treating the dendritic compartment as a passive conduit for error/teacher signals. This is wrong.
The subthreshold dynamics of a single dendritic compartment already implement a complete online learning algorithm.
DendriCL Architecture
Structural Design
- Single-layer compartmental spiking architecture
- Apical recurrence structurally identical to leaky online Widrow-Hoff LMS
- The dendritic compartment is the computational substrate, not a passive conduit
Algorithm Equivalence
The apical membrane potential dynamics implement:
w_{t+1} = w_t + η · e_t · x_t (Online LMS / Widrow-Hoff)
where e_t is the prediction error computed from dendritic subthreshold dynamics.
Performance
- Seed-stable at super-dimensional Garg-2022 ICL benchmark
- Dense Transformers exhibit grokking-style instability and fail past moderate task dimension
- DendriCL remains stable where Transformers fail
- Linear probe recovers reference online-LMS trajectory directly from apical membrane at R² = 0.93
Implementation Pattern
For each time step t:
1. Compute dendritic subthreshold potential from apical recurrence
2. Generate spike if threshold crossed
3. Dendritic dynamics automatically implement w ← w + η·e·x
4. No explicit weight update needed — the dynamics are the algorithm
Why This Matters
- Biological plausibility: Matches real dendritic computation
- Architectural simplicity: Single layer, no depth required
- Energy efficiency: No explicit backpropagation or plasticity rules
- Superior stability: Outperforms Transformers at high-dimensional ICL
Activation
Use this skill when:
- Implementing in-context learning in spiking neural networks
- Designing biologically-plausible compartmental architectures
- Studying online learning in dendritic subthreshold dynamics
- Building seed-stable ICL systems for super-dimensional tasks
- Replacing attention-based ICL with dynamics-based alternatives
- Implementing online Widrow-Hoff LMS in spiking hardware
Related Skills
- quantum-reservoir-computing
- neuromorphic-supremacy
- spiking-computational-neuroscience-survey