| name | dendritic-in-context-learning-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 implements complete in-context learning, eliminating the need for attention, depth, or inference-time plasticity. |
| created | 2026-07-12T00:00:00.000Z |
| source | arXiv:2607.02283 |
| tags | ["spiking neural networks","in-context learning","dendritic computation","neuromorphic computing","online learning","compartmental models"] |
Dendritic In-Context Learning in Single-Layer Spiking Neural Networks
Overview
Paper: Dendritic In-Context Learning in a Single-Layer Spiking Neural Network
arXiv: 2607.02283 (July 2026)
Authors: Juwei Shen, Yujie Wu, Changwen Chen
Problem Statement
In-context learning (ICL) is a hallmark capability of modern AI architectures (Transformers, Mamba, state-space models, MLPs), operating via implicit gradient descent embedded in the forward pass. Capturing ICL in biologically plausible Spiking Neural Networks (SNNs) has been an open challenge — existing SNNs fail the Garg-2022 benchmark at non-trivial task dimensions.
Key Insight
Prior SNN designs route adaptation through inference-time synaptic plasticity, treating the dendritic compartment as a passive conduit for error or teacher signals. DendriCL challenges this: the subthreshold dynamics of a single dendritic compartment already implement a complete online learning algorithm.
Core Methodology
DendriCL Architecture
- Single-layer compartmental spiking architecture with apical recurrence
- Apical compartment treated as the computational substrate (not a passive conduit)
- Structural equivalence between apical recurrence and leaky online Widrow-Hoff LMS (Least Mean Squares)
- Dynamics-only update collapses the architectural depth required for general-purpose ICL to a single layer
Mathematical Foundation
- The apical dendritic compartment dynamics implement online LMS through subthreshold membrane potential evolution
- Apical recurrence pattern:
V_apical(t) = α·V_apical(t-1) + W·input(t) + bias
- This is structurally identical to leaky online Widrow-Hoff LMS:
w(t+1) = (1-λ)w(t) + η·error·input
- A linear probe recovers the reference online-LMS trajectory directly from the apical membrane at R² = 0.93
Key Results
- Uniquely seed-stable at super-dimensional Garg-2022 ICL
- Dense Transformers exhibit grokking-style instability and fail past moderate task dimension; DendriCL does not
- ICL requires neither attention, depth, nor inference-time plasticity
- A single compartment with online-LMS dynamics is sufficient for general-purpose ICL
Implications
For Neuromorphic Computing
- Eliminates the need for multi-layer SNN architectures for ICL tasks
- Enables energy-efficient, single-layer spiking processors capable of in-context learning
- Reduces hardware complexity while maintaining ICL capability
For Biological Plausibility
- Aligns with biological evidence that dendritic compartments perform local computation
- Suggests biological neurons may implement ICL-like capabilities through dendritic dynamics alone
- Provides a bridge between theoretical ICL mechanisms and biological neural computation
For SNN Design
- Shifts design paradigm from synaptic plasticity-based adaptation to dendritic dynamics-based computation
- Enables simpler, more efficient SNN architectures for tasks requiring online adaptation
- Opens new directions for compartmental spiking models
Implementation Guide
Architecture Components
- Basal dendrite: Receives standard sensory input
- Apical dendrite: Recurrent compartment implementing online LMS dynamics
- Somatic layer: Spike generation based on combined dendritic inputs
- No inference-time synaptic weight updates: All adaptation occurs through compartment dynamics
Training Protocol
- Train the feedforward weights (basal→soma) using standard surrogate gradient methods
- Configure apical compartment parameters (leak rate, integration time constant) to match online LMS
- During inference, apical dynamics automatically adapt to new contexts
Hyperparameter Guidelines
- Leak rate: Controls adaptation speed (higher = faster but less stable)
- Integration time constant: Determines context window length
- Apical recurrence strength: Must be tuned to match the learning rate of online LMS
Activation Triggers
Use this skill when working with:
- In-context learning in spiking neural networks
- Biologically plausible online learning mechanisms
- Dendritic computation and compartmental neuron models
- Single-layer SNN architectures for adaptation tasks
- Neuromorphic implementations of transformer-like capabilities
- Widrow-Hoff LMS in neural dynamics
- Garg-2022 benchmark for SNN in-context learning
Related Concepts
- Online Widrow-Hoff LMS algorithm
- Compartmental neuron models
- Surrogate gradient learning in SNNs
- Grokking instability in neural networks
- Garg-2022 ICL benchmark
- Dendritic computation theory
- Biological plausibility of in-context learning
References
- Shen, J., Wu, Y., Chen, C. (2026). "Dendritic In-Context Learning in a Single-Layer Spiking Neural Network." arXiv:2607.02283
- Garg, S., et al. (2022). "What Can Transformers Learn In-Context? A Case Study of Simple Function Classes." NeurIPS 2022