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

A research team from institutions including the University of Stuttgart and ETH Zurich has published a paper on DendroNN, a novel neural network architecture that mimics the computational power of biological dendrites. The core innovation is a sequence detection mechanism that identifies unique incoming spike patterns as spatiotemporal features, allowing the network to process event-based data—like signals from neuromorphic cameras or audio sensors—with high temporal accuracy without resorting to computationally expensive recurrence or explicit delays.

The model is trained through a unique 'rewiring' phase instead of standard backpropagation. This gradient-free process allows the network to memorize frequently occurring spike sequences and prune non-discriminative ones, directly addressing the challenge of training non-differentiable spike-based systems. This approach maintains competitive classification accuracy across various event-based time series benchmarks.

Crucially, the team also designed a corresponding asynchronous digital hardware architecture. It uses a 'time-wheel' mechanism and leverages the model's inherent dynamic/static sparsity and quantization. This event-driven design eliminates the need for per-step global updates common in other neuromorphic models. The result is a system that demonstrated up to 4x higher hardware efficiency (presumably in operations per joule) compared to state-of-the-art neuromorphic hardware while maintaining comparable accuracy on an audio classification task, paving the way for ultra-low-power embedded AI.