| name | dendricl-icl-single-layer-snn |
| description | DendriCL methodology for dendritic in-context learning in single-layer spiking neural networks. The apical compartment's subthreshold dynamics implement online Widrow-Hoff LMS, enabling general-purpose ICL without attention, depth, or inference-time plasticity. First SNN to solve Garg-2022 ICL benchmark at d≥30 where Transformers fail. Trigger: dendritic computation, in-context learning SNN, compartmental neuron, online LMS, biological ICL, neuromorphic ICL, apical dendrite, single-layer learning
|
| arxiv_id | 2607.02283v1 |
| date | 2026-07-02 |
| authors | Juwei Shen, Yujie Wu, Changwen Chen (HK PolyU) |
| categories | ["cs.NE","cs.LG"] |
| tags | ["spiking neural network","in-context learning","dendritic computation","compartmental neuron","online LMS","neuromorphic"] |
DendriCL: Dendritic In-Context Learning in a Single-Layer Spiking Neural Network
Core Insight
In-context learning (ICL) does NOT require attention, depth, or inference-time plasticity. A single dendritic compartment with online-LMS dynamics is sufficient.
The subthreshold dynamics of a single apical dendritic compartment implement a complete online learning algorithm (leaky Widrow-Hoff LMS). By treating the compartment as the computational substrate rather than a passive conduit, ICL emerges from dynamics alone with frozen synaptic weights at inference.
Architecture
Single-layer compartmental spiking network with d_model = 384 parallel pyramidal-like units:
Per-unit recurrence at context position t:
uB(t) = WB * xt (basal projection)
ŷt = uA(t)^T * WA * xt (scalar prediction)
et = (1 - flagt) * (yt - ŷt) (gated error)
uA(t+1) = α * uA(t) + γ * et * WA * xt (apical online-LMS)
vsoma(t) = gB * uB(t) + gA * WA,out * uA(t) (somatic integration)
s(t) = ⊮[vsoma(t) > θ] (LIF spike, soft reset)
Key design choices:
- Apical state uA is NOT reset by somatic spikes — evolves continuously across full context
- Error et is gated OFF at query position
- All synaptic weights frozen at inference time
- ~0.75M total parameters, single layer
Structural Equivalence: Apical ≡ Leaky Online LMS
When WA = I, the apical update reduces to:
ŵ_{t+1} = α * ŵ_t + γ * (yt - ŵ_t^T * xt) * xt
This is classical leaky Widrow-Hoff LMS. Under i.i.d. inputs and linear targets, E||uA(k) - w||² = O(d/k) — the classical LMS convergence theorem.
Linear probe recovers reference online-LMS trajectory from apical membrane at R² = 0.93, confirming the algorithm is structurally embedded in the dynamics.
Key Results
Garg-2022 ICL Benchmark
- d=10: DendriCL R² = 0.807 (competitive with all architectures)
- d=20: DendriCL R² = 0.820 (best spike-based model, within 3pp of non-spiking ablation)
- d=30: DendriCL R² = 0.807 (ONLY architecture maintaining >0.5)
- d=40: DendriCL R² = 0.787 (Transformers collapse to chance)
- d=50: DendriCL R² = 0.649 (all others at chance floor)
Seed Stability
- DendriCL: σ ≤ 0.036 across entire super-d range (uniquely stable)
- Transformer: bimodal at d=30, collapses at d≥40 (architectural failure, not budget)
- Spikformer: degrades gracefully but trails by 0.17 at d=30
Efficiency
- ~4× spike reduction over Pure LIF
- Projected ~10× Loihi-class energy advantage
- Architectural simplicity and inference-time efficiency co-vary (not trade off)
Biological Grounding
- Three-compartment layout (apical-basal-soma) matches cortical layer-5 pyramidal neurons
- Apical compartment carries persistent multi-dimensional subthreshold voltage with calcium plateaus on 100+ ms timescales
- Consistent with predictive coding framework (apical = error-driven dynamics)
- Falsifiable hypothesis: in vivo recordings should reveal LMS-like dynamics in apical tufts
Contrast with Prior Work
| Model | Apical Role | Adaptation Mechanism |
|---|
| Urbanczik-Senn 2014 | Teacher signal | Plasticity-driven |
| Sacramento 2018 | Backprop error | Plasticity-driven |
| Iyer 2022 (Active Dendrites) | External context gate | Plasticity-driven |
| Miconi 2018 | Differentiable plasticity | Hebbian fast weights |
| DendriCL (Ours) | Online LMS estimator | Dynamics-driven, frozen weights |
Practical Applications
- Neuromorphic ICL: Deploy in-context learning on Loihi/SpiNNaker without inference-time weight updates
- Energy-efficient adaptation: Single-layer architecture enables ~10× energy savings
- Brain-inspired AI: Demonstrates that biological dendritic computation is computationally sufficient for ICL
- Robust high-dimensional learning: Only architecture seed-stable at d≥30
- Hardware-software co-design: Compartmental dynamics map naturally to analog neuromorphic circuits
Training Details
- BPTT from scratch, AdamW, lr=1e-3, weight decay 1e-4, cosine schedule
- Batch size 64, 10k steps (baseline) or 50k steps (compute-matched)
- LIF surrogate gradients: arctan approximation
- 3 seeds per config (5 seeds for d∈{20,30})
- No pre-training, no plasticity at inference
Implications
- ICL is a generic property of trained sequence models — now extended to spiking substrate
- The "depth hypothesis" for ICL is wrong: a single compartment suffices
- Dendritic computation is not just biologically plausible — it's computationally optimal for ICL
- Opens path to fully neuromorphic in-context learning systems