| name | dendritic-icl-snn |
| description | Dendritic In-Context Learning (DendriCL) — single-layer compartmental SNN that achieves ICL via dendritic subthreshold dynamics implementing online LMS. Use when: designing SNN architectures for in-context learning, exploring biologically-plausible ICL mechanisms, studying dendritic computation, implementing seed-stable ICL beyond moderate task dimensions, analyzing compartmental neuron models. |
| tags | ["spiking-neural-networks","in-context-learning","dendritic-computation","neuromorphic","online-lms","compartmental-snn","biological-plausibility","icl-benchmark","seed-stability","single-layer-snn","garg-2022-benchmark"] |
| published | "2026-07-02T00:00:00.000Z" |
| arxiv_id | 2607.02283 |
| authors | ["Juwei Shen","Yujie Wu","Changwen Chen"] |
Dendritic In-Context Learning (DendriCL)
arXiv: 2607.02283 (2026-07-02)
Authors: Juwei Shen, Yujie Wu, Changwen Chen
Overview
In-context learning (ICL) has been demonstrated in Transformers, Mamba, state-space models, and MLPs through implicit gradient descent embedded in the forward pass. However, existing Spiking Neural Networks fail the Garg-2022 ICL 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/teacher signals. The authors challenge this assumption — the subthreshold dynamics of a single dendritic compartment already implement a complete online learning algorithm.
DendriCL is a single-layer compartmental spiking architecture whose apical recurrence is structurally identical to leaky online Widrow-Hoff LMS. This dynamics-only update collapses the architectural depth required for general-purpose ICL to a single layer.
Core Technical Contributions
1. Dendritic Compartment as Computational Substrate
The apical membrane potential of the dendritic compartment evolves as:
$$v[t] = (1 - \lambda) \cdot v[t-1] + \lambda \cdot (w^T x[t]) - \eta \cdot e[t]$$
where:
- $\lambda$ is the leaky factor controlling temporal integration
- $w$ are synaptic weights
- $x[t]$ are input spikes
- $\eta$ is the learning rate
- $e[t]$ is the prediction error signal
This directly implements leaky online Widrow-Hoff (LMS) learning without backpropagation or inference-time plasticity rules.
2. Architectural Design
- Single-layer compartmental neurons with separate basal and apical dendritic compartments
- Apical recurrence creates the feedback loop for online LMS
- Basal compartment receives input spikes and produces output
- Apical compartment implements the learning dynamics
- No depth, no attention, no inference-time plasticity required
3. Seed-Stable ICL at Super-Dimensional Dimensions
Key empirical finding: DendriCL is uniquely seed-stable at super-dimensional Garg-2022 ICL, where dense Transformers exhibit grokking-style instability and fail past moderate task dimension.
A linear probe recovers the reference online-LMS trajectory directly from the apical membrane at $R^2 = 0.93$, confirming the algorithm is structurally embedded in the dynamics rather than implicitly discovered during training.
4. Key Findings
- ICL requires neither attention, depth, nor inference-time plasticity
- A single compartment with online-LMS dynamics is sufficient
- Seed stability: DendriCL maintains performance across random seeds at task dimensions where Transformers fail
- The online-LMS trajectory is directly recoverable from apical membrane potential ($R^2 = 0.93$)
Implementation Guidelines
Compartmental Neuron Model
Each neuron has two compartments:
Basal compartment (input integration):
- Receives input spikes x[t]
- Integrates with synaptic weights w
- Produces output y[t]
Apical compartment (learning dynamics):
- Receives prediction error e[t]
- Updates internal state via leaky LMS
- Provides feedback to basal compartment
Online LMS Implementation
v_apical[t] = (1 - leak) * v_apical[t-1] + leak * (w.T @ x[t])
prediction = sigmoid(v_apical[t])
error = target[t] - prediction
v_apical[t] += learning_rate * error
Garg-2022 ICL Benchmark
DendriCL should be evaluated on:
- Garg-2022 ICL benchmark with varying task dimensions
- Seed stability analysis across multiple random initializations
- Linear probe analysis to verify LMS trajectory recovery
Comparison Baselines
Compare against:
- Dense Transformers (grokking-style failure at high dimensions)
- Prior SNN-ICL approaches (fail at non-trivial task dimensions)
- Reference online-LMS trajectory
Design Patterns
Pattern 1: Single-Layer ICL via Compartmental Dynamics
Replace multi-layer SNN architectures with single-layer compartmental neurons for ICL tasks. The apical dendritic compartment naturally implements the learning algorithm.
Pattern 2: Seed-Stable Architecture Design
When seed instability is observed at high task dimensions, consider whether the learning dynamics are structurally embedded (like DendriCL) or implicitly discovered (like Transformers).
Pattern 3: Linear Probe Verification
Use linear probes on internal states to verify that target algorithms (e.g., LMS) are structurally embedded in the dynamics.
Activation Keywords
dendritic, dendrite, in-context learning, ICL, compartmental, online-LMS, Widrow-Hoff, seed-stable, Garg-2022, single-layer, spiking, SNN, apical, basal, subthreshold, implicit gradient, biologically plausible, grokking, stability
Pitfalls
-
Prior SNN-ICL failures: Existing SNNs fail Garg-2022 at non-trivial dimensions. Don't assume SNNs can do ICL without compartmental dynamics.
-
Synaptic plasticity ≠ ICL: Prior approaches using inference-time synaptic plasticity fail. The learning must be embedded in compartmental dynamics.
-
Seed stability matters: Transformers exhibit grokking-style instability at high dimensions. DendriCL's seed stability is a key advantage.
-
Apical vs basal compartments: Must maintain separate compartments. Treating dendrites as passive conduits loses the ICL capability.
-
Garg-2022 benchmark: Standard ICL benchmarks may not capture seed stability. Use Garg-2022 for rigorous evaluation.
Related Skills
spiking-neural-network-analysis
spikingjelly-framework
adaptive-spiking-neuron-asn
meta-learning-in-context-brain-decoding
dendrocentric-snn-event-classification
working-memory-heterogeneous-delays