| name | dendritic-icl-snn |
| description | DendriCL methodology for dendritic in-context learning in single-layer spiking neural networks. Demonstrates that a single dendritic compartment with online-LMS dynamics is sufficient for general-purpose ICL without attention, depth, or inference-time plasticity. |
| tags | ["spiking-neural-network","in-context-learning","dendritic-computation","neuromorphic","biological-plausibility"] |
| arxiv_id | 2607.02283 |
| authors | ["Juwei Shen","Yujie Wu","Changwen Chen"] |
| published | 2026-07-02 |
Dendritic In-Context Learning in Single-Layer SNN (DendriCL)
Core Insight
In-context learning requires neither attention, depth, nor inference-time plasticity: a single compartment with online-LMS dynamics is sufficient.
This paper challenges the fundamental assumption that ICL in SNNs requires synaptic plasticity during inference. Instead, it demonstrates that the subthreshold dynamics of a single dendritic compartment already implement a complete online learning algorithm.
Key Contributions
1. Paradigm Shift: Dendrite as Computational Substrate
- Previous assumption: Dendrites are passive conduits for error/teacher signals
- New insight: Dendritic compartment IS the computational substrate
- Subthreshold dynamics of single dendritic compartment = complete online learning algorithm
2. DendriCL Architecture
- Single-layer compartmental spiking architecture
- Apical recurrence structurally identical to leaky online Widrow-Hoff LMS
- Dynamics-only update collapses architectural depth to single layer
- No attention mechanism required
- No inference-time synaptic plasticity needed
3. Performance Breakthrough
- First SNN to pass Garg-2022 ICL benchmark at non-trivial task dimensions
- Seed-stable at super-dimensional ICL tasks
- Dense Transformers exhibit grokking-style instability at same dimension; DendriCL remains stable
- Linear probe recovers reference online-LMS trajectory from apical membrane with R² = 0.93
- Algorithm is structurally embedded in dynamics, not implicitly discovered during training
Technical Details
Architecture Components
Single SNN Layer
├── Somatic compartment (standard LIF dynamics)
└── Apical dendritic compartment
├── Recurrent connections
└── Subthreshold dynamics ≡ online Widrow-Hoff LMS
Mathematical Equivalence
The apical dendritic recurrence implements:
- Online LMS (Least Mean Squares) update rule
- Leaky integration of prediction errors
- Real-time adaptation without explicit gradient computation
Biological Plausibility
- Single compartment with realistic biophysics
- No backpropagation through time
- No separate learning phase vs inference phase
- Consistent with known dendritic computation in pyramidal neurons
Implications
For Neuroscience
- Provides computational theory for dendritic function
- Explains how single neurons might perform complex learning
- Supports "dendritic democracy" hypothesis
- Bridges gap between biophysics and machine learning
For AI/ML
- Dramatically reduces architectural complexity for ICL
- Eliminates need for deep stacks or attention
- Enables ultra-efficient neuromorphic implementations
- Suggests new directions for brain-inspired AI
For Neuromorphic Computing
- Single-layer implementation = minimal hardware
- Event-driven computation = energy efficient
- Biologically plausible = potential for brain-like chips
Experimental Validation
Garg-2022 Benchmark
- Standard test for ICL capability
- Tests ability to learn new tasks from context examples
- Previous SNNs failed at non-trivial dimensions
- DendriCL succeeds where Transformers fail (super-dimensional regime)
Mechanistic Verification
- Linear probe on apical membrane potential
- Recovers reference LMS trajectory with R² = 0.93
- Confirms algorithm is embedded in dynamics, not learned implicitly
Implementation Guidelines
When to Use
- Tasks requiring in-context learning
- Energy-constrained neuromorphic deployment
- Biologically plausible neural modeling
- Single-layer architectures preferred
Design Principles
- Dendrite-first: Treat dendritic compartment as primary computational unit
- Dynamics-as-algorithm: Let subthreshold dynamics implement learning
- Minimal architecture: Single layer sufficient for ICL
- No plasticity switching: Continuous adaptation through dynamics
Integration with Existing Systems
- Can replace deep Transformer stacks for ICL tasks
- Compatible with existing SNN training frameworks
- Drop-in replacement for attention-based ICL modules
Research Directions
Open Questions
- How does this scale to multi-task learning?
- Can multiple dendritic compartments enable compositionality?
- What is the capacity limit of single-compartment ICL?
- How does this relate to working memory in prefrontal cortex?
Extensions
- Multi-compartment DendriCL for hierarchical processing
- Hybrid DendriCL-Transformer architectures
- Hardware implementation on neuromorphic chips
- Application to real-time adaptive control
Citation
@article{shen2026dendritic,
title={Dendritic In-Context Learning in a Single-Layer Spiking Neural Network},
author={Shen, Juwei and Wu, Yujie and Chen, Changwen},
journal={arXiv preprint},
year={2026},
eprint={2607.02283},
archivePrefix={arXiv},
primaryClass={cs.NE}
}
Related Work
- Transformers for ICL: Attention-based in-context learning
- Mamba/SSMs: State-space models for sequence learning
- Dendritic computation: Previous work on dendritic processing
- SNN learning rules: STDP, surrogate gradients, etc.
- Biological plausibility: Brain-inspired AI approaches
Activation Triggers
Use this skill when working on:
- In-context learning in spiking networks
- Dendritic computation models
- Single-layer neural architectures
- Biologically plausible learning
- Neuromorphic computing design
- Energy-efficient AI
- Brain-inspired algorithms
Keywords: dendritic, in-context learning, ICL, spiking neural network, SNN, single-layer, online learning, Widrow-Hoff, LMS, biological plausibility, neuromorphic