[Paper] DendroNN: Dendrocentric Neural Networks for Energy-Efficient Classification of Event-Based Data

Source: arXiv - 2603.09274v1

Overview

The paper introduces DendroNN, a new class of spiking neural networks that mimic the way biological dendrites detect specific spike sequences. By turning dendritic sequence detection into a trainable, event‑driven architecture, the authors achieve high‑accuracy classification of event‑based data while dramatically cutting energy consumption—making it a promising candidate for low‑power neuromorphic hardware.

Key Contributions

• Dendrocentric network design: A novel spiking architecture that treats dendritic branches as sequence detectors, turning temporal spike patterns into discriminative features.
• Gradient‑free rewiring training: A biologically inspired “rewiring” phase that learns which spike sequences to keep or discard, enabling training without back‑propagation through non‑differentiable spikes.
• Dynamic/static sparsity exploitation: The network naturally prunes unused dendritic branches, yielding both static (structural) and dynamic (activation‑time) sparsity.
• Asynchronous digital hardware prototype: Introduces a “time‑wheel” event‑driven processor that eliminates per‑step global updates, a common bottleneck in recurrent or delay‑based SNNs.
• Energy‑efficiency results: Demonstrates up to 4× lower energy per inference compared to state‑of‑the‑art neuromorphic platforms on audio event classification, with comparable accuracy.

Methodology

1. Dendritic Sequence Detection: Each dendritic branch monitors incoming spikes and fires only when a predefined temporal order (e.g., spike A → spike B → spike C within a time window) occurs. This creates a set of spatiotemporal “motifs” that act as high‑level features.
2. Rewiring Phase:
   • Memorization: During an unsupervised exposure to training data, the network records frequently observed spike sequences.
   • Pruning: Sequences that never contribute to correct class decisions are removed, reducing the number of active dendrites.
   • This process is akin to synaptic growth and elimination in biology and sidesteps the need for gradient descent on discrete spike events.
3. Network Architecture: A shallow feed‑forward SNN where the first layer consists of dendritic detectors, followed by a simple read‑out layer that aggregates the binary outputs into class scores.
4. Hardware Implementation: The authors design an asynchronous digital accelerator that uses a time‑wheel—a rotating pointer that timestamps incoming events—so that each dendrite updates only when its specific sequence pattern is matched, avoiding global clock cycles.

Results & Findings

• Benchmark Datasets: Tested on several event‑based time‑series benchmarks (e.g., N‑MNIST, DVS‑Gesture, audio spike‑encoded speech).
• Accuracy: Achieved classification scores within 1–3 % of the best recurrent or delay‑based SNNs, despite using a shallower, feed‑forward topology.
• Energy Consumption: On an audio classification task, the DendroNN hardware consumed ~0.25 nJ per inference, roughly 4× less than leading neuromorphic chips (e.g., Loihi, TrueNorth) while delivering similar accuracy.
• Sparsity Metrics: Static sparsity (pruned dendrites) reached ~70 % reduction in parameters; dynamic sparsity (event‑driven updates) yielded ~85 % fewer clock cycles per inference.

Practical Implications

• Edge AI Devices: DendroNN’s event‑driven nature makes it ideal for battery‑powered sensors (audio wake‑words, event‑based cameras) where every microjoule counts.
• Neuromorphic Accelerators: The time‑wheel architecture can be integrated into existing digital ASIC flows, offering a drop‑in replacement for recurrent SNN blocks without redesigning the whole pipeline.
• Low‑Latency Processing: Because updates occur only on relevant spike patterns, inference latency scales with the actual information content, not with a fixed time step—beneficial for real‑time detection (e.g., gesture recognition).
• Simplified Training Pipelines: The rewiring approach removes the need for surrogate gradient tricks, allowing developers to train models with standard event‑stream data pipelines and simple rule‑based pruning scripts.

Limitations & Future Work

• Sequence Length Constraints: The current dendritic detectors are limited to relatively short spike patterns; longer temporal dependencies may still require recurrence or external memory.
• Hardware Prototyping Scope: The hardware evaluation is based on a digital ASIC simulation; silicon validation and scaling to larger networks remain open.
• Generalization to Vision: While audio and simple event‑based datasets are covered, applying DendroNN to high‑resolution event‑camera streams may demand more sophisticated dendritic encoding schemes.
• Training Overhead: The rewiring phase can be computationally intensive for very large datasets, suggesting a need for more efficient online pruning algorithms.

Overall, DendroNN opens a fresh pathway for energy‑efficient spatiotemporal AI, bridging neuroscience insights with practical neuromorphic engineering.

Authors

• Jann Krausse
• Zhe Su
• Kyrus Mama
• Maryada
• Klaus Knobloch
• Giacomo Indiveri
• Jürgen Becker

Paper Information

• arXiv ID: 2603.09274v1
• Categories: cs.LG, cs.AI, cs.AR, cs.ET, cs.NE
• Published: March 10, 2026
• PDF: Download PDF