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

This paper introduces DendroNN, a novel dendrocentric neural network that leverages non-differentiable sequence detection and a rewiring phase to efficiently classify event-based spatiotemporal data, achieving competitive accuracy with up to 4x higher energy efficiency than state-of-the-art neuromorphic hardware through a dedicated asynchronous digital architecture.

The Gist

Imagine your brain is a bustling city. In a traditional computer, information flows like a traffic light system: everything moves in strict, synchronized steps, checking every intersection at the exact same time, whether a car is there or not. This is how most Artificial Intelligence works today. It's reliable, but it's also wasteful, like leaving all the streetlights on even when the road is empty.

DendroNN is a new kind of AI that tries to work more like a real human brain, specifically focusing on how our neurons "listen" to the world. Here is the story of how it works, broken down into simple concepts.

1. The Problem: The "Waiting Room" vs. The "Detective"

Most current AI models (called Spiking Neural Networks) are like a waiting room. They collect all the messages (spikes) that arrive and just count them. If enough people show up, they make a decision. But they don't care when the people arrived or in what order.

• The Flaw: If you say "Cat" and "Dog" to a waiting room, it just hears two words. It doesn't know the difference between "Cat then Dog" and "Dog then Cat." To fix this, old AI models try to remember the past by keeping a giant notebook of everything that happened, which takes up a lot of energy and space.

2. The Solution: The "Dendritic Detective"

In our biological brains, neurons have tree-like branches called dendrites. Scientists recently realized these aren't just wires; they are tiny detectives.

A dendrite doesn't just count spikes; it looks for a specific pattern. It's like a bouncer at a club who only lets people in if they arrive in a specific order: First the guy in the red hat, then the woman with the blue scarf, exactly 2 seconds later. If the order is wrong, or the timing is off, the bouncer says, "No entry."

DendroNN copies this. Instead of a waiting room, it builds a network of these "dendritic detectives." Each detective is trained to listen for one specific sequence of events.

3. How It Learns: The "Rewiring" Phase

Here is the tricky part: You can't teach these detectives using the usual math methods (gradients) because their rules are too rigid (it's either "yes, that's the pattern" or "no, it's not").

So, the researchers invented a Rewiring Phase. Imagine you have a room full of detectives, but they are all guessing randomly.

1. The Trial: You show them thousands of examples (like audio clips or Morse code).
2. The Scorecard: If a detective hears a pattern that happens often, they get a "point." If they hear a pattern that never happens, they get a "penalty."
3. The Shuffle: The detectives with low scores get fired and replaced with new ones who are assigned random new patterns to look for.
4. The Freeze: The detectives with high scores get "frozen" in place. They keep their specific pattern and become the permanent experts.

This way, the network automatically figures out which patterns are important without needing complex math. It's like evolution: the fittest patterns survive.

4. The Hardware: The "Time Wheel"

To make this super fast and energy-efficient, the researchers built a special hardware chip.

• Old Way (The Conveyor Belt): Traditional chips process time like a conveyor belt. Every second, they have to shift every single item on the belt one step forward to make room for the new one. This takes a lot of energy, even if the belt is empty.
• DendroNN Way (The Time Wheel): Imagine a giant clock face (a wheel). Instead of moving the items, the "hand" of the clock moves. When a detective needs to check if a message arrived 5 seconds ago, the system just looks at the spot on the wheel where the hand was 5 ticks ago.
  • The Magic: If no one sent a message, the system does nothing. It doesn't waste energy checking empty spots. It only wakes up when an event actually happens.

5. The Results: Why It Matters

The paper tested this on three things:

• Morse Code: Decoding dots and dashes.
• Handwritten Digits: Reading numbers written as a stream of pixels.
• Audio: Recognizing spoken digits.

The Winner:
DendroNN was just as accurate as the best existing AI models but used 4 times less energy.

• Analogy: If a standard AI chip is a gas-guzzling truck that idles while waiting for traffic, DendroNN is a hybrid car that only uses fuel when you actually press the gas pedal.

Summary

DendroNN is a new type of AI that:

1. Listens for patterns (like a detective) instead of just counting.
2. Learns by trial and error (rewiring itself) to find the best patterns.
3. Runs on a "Time Wheel" hardware that saves massive amounts of energy by only working when something actually happens.

It's a step toward making smart devices that are as efficient and fast as our own brains, perfect for things like hearing aids, self-driving cars, or sensors that need to run for years on a single battery.

Technical Summary

Here is a detailed technical summary of the paper "DendroNN: Dendrocentric Neural Networks for Energy-Efficient Classification of Event-Based Data."

1. Problem Statement

Temporal sequences are fundamental to neural computation, yet current artificial neural systems struggle to decode them efficiently under constraints of locality, sparsity, and low power.

• Limitations of Traditional Models: Standard Deep Neural Networks (DNNs) treat time as independent static frames, ignoring causality. Recurrent Neural Networks (RNNs) and Spiking Neural Networks (SNNs) attempt to address this but often rely on recurrence or delay buffers.
• Hardware Inefficiency:
  • RNNs/Recurrent SNNs: Require global state updates for every time step, leading to high energy and area costs that scale with sequence length.
  • Delay-based SNNs: Achieve high accuracy by aligning signal timing via delay buffers but incur massive hardware overhead (expensive shifting memory in digital or large capacitors in analog).
  • Standard SNNs: Basic units integrate inputs commutatively (summing spikes regardless of order), failing to detect specific spatiotemporal patterns without external timing mechanisms.
• The Gap: There is a need for a model that mimics biological dendritic sequence detection (where specific spike orders trigger firing) but is compatible with event-driven, asynchronous digital hardware without requiring global synchronization or heavy memory access.

2. Methodology: The DendroNN Model

The authors propose DendroNN, a neural network architecture that abstracts biological dendritic branches as compartmentalized processing units capable of detecting structured spike sequences.

A. Core Mechanism: Sequence Detection

• Spatiotemporal Features: A DendroNN unit is tuned to a specific sequence defined by:
  1. Spatial Origin: Which presynaptic neurons ($X$) fire.
  2. Temporal Order: The permutation ($\sigma$) of the spikes.
  3. Inter-Spike Intervals (ISIs): The precise time gaps ($\Delta t_i$) between spikes.
• Activation Logic: A unit emits an output spike only if inputs arrive from the correct spatial origins in the correct temporal order with the correct ISIs.
  • Mathematically, this is an AND operation across time and space: $o_{t_0} = \bigwedge I_{t_i}(\tilde{x}_i)$.
  • Deviations in order or timing suppress activation.
• Parallelism: A single unit can detect multiple instances of the same sequence in parallel, filtering out noise and irrelevant spikes.

B. Training Strategy: The Rewiring Phase

Since the sequence-defining parameters (integer ISIs and binary connections) are non-differentiable, standard backpropagation cannot be used to train the hidden layer. The authors introduce a gradient-free rewiring phase:

1. Longevity State: During inference on training data, units track a "longevity" score based on how often their current random sequence appears in the data.
2. Rewiring: Units with low longevity (rare sequences) are rewired (randomly redrawn connections and ISIs). Units with high longevity are "frozen."
3. Selectivity Check: To ensure the frozen sequences are discriminative, a validation step checks if a unit's firing is selective to specific classes. Only units with high class selectivity are permanently retained.
4. Supervised Fine-tuning: Once the hidden layer is fixed, the final linear output layer is trained using standard backpropagation.

C. Hardware Architecture: Asynchronous Time-Wheel

To support DendroNN on hardware, the authors designed a near-memory asynchronous digital architecture:

• Event-Driven: No global clock; operations occur only when spikes arrive.
• Time-Wheel Mechanism: Instead of shifting delay buffers (which is memory-bandwidth intensive), the architecture uses a circular pointer ($p$) and a "time-wheel" to track inter-spike intervals.
  • Packed Two-Generation Slots: To handle sequences longer than the wheel size, each slot stores two bits (current and next generation) to avoid wrap-around ambiguity.
• Sparsity: The architecture exploits the high dynamic and static sparsity of DendroNNs, accessing memory only for active units.

3. Key Contributions

1. Dendrocentric Neural Network (DendroNN): A novel architecture that abstracts dendritic sequence detection into a binary, event-driven model capable of identifying unique spatiotemporal patterns.
2. Gradient-Free Rewiring: A training methodology that autonomously discovers relevant temporal structures in non-differentiable spike sequences without requiring surrogate gradients.
3. Asynchronous Hardware Co-Design: A custom digital hardware architecture using a time-wheel mechanism that eliminates per-step global updates, scaling memory access with event sparsity rather than time window length.
4. End-to-End System: A complete pipeline from algorithm design, training, compression (pruning/quantization), to hardware implementation and evaluation.

4. Results

The DendroNN was evaluated on four datasets and compared against state-of-the-art SNNs and neuromorphic hardware.

A. Algorithmic Performance

• NeuroMorse (Morse Code): Achieved 90.22% accuracy on noise-free data (vs. ~12% for standard SNNs) and maintained >45% accuracy under high noise. It demonstrated superior "uncertainty detection" by remaining silent on null-class samples.
• Sequential MNIST (sMNIST & p-sMNIST): Achieved 97.50% (sMNIST) and 94.88% (p-sMNIST), competing with top SNNs while using significantly fewer parameters.
• Spiking Heidelberg Digits (SHD): Achieved 89.24% accuracy. While slightly lower than delay-based SNNs (which reach ~93%), it significantly outperformed recurrent SNNs and feed-forward SNNs.
• Memory Footprint: DendroNNs consistently required 10–100x less memory than comparable delay-based or recurrent models due to binary weights and sparse connections.

B. Hardware Efficiency (SHD Task)

Implemented in GlobalFoundries 22FDX FDSOI technology and compared against Loihi2, DenRAM, ReckOn, and ElfCore:

• Energy Efficiency: DendroNN achieved up to 4.1x higher energy efficiency than delay-based hardware and 2.8x higher than recurrence-based hardware.
• Throughput: Up to 1.9x higher throughput than delay-based hardware.
• Area Efficiency: Over 2x higher area efficiency than recurrence-based hardware due to the lack of complex state storage (membrane potentials) and binary weights.
• Mechanism: The efficiency gains stem from eliminating low-arithmetic-intensity vector operations and memory read-modify-write cycles typical of other SNN hardware.

5. Significance and Conclusion

• Bridging Biology and Hardware: DendroNN successfully translates biological dendritic computation (sequence sensitivity) into a format that is natively efficient for digital, asynchronous hardware.
• Scalability: The approach demonstrates that sparse, event-driven systems can achieve competitive accuracy on complex temporal tasks while drastically reducing the memory and energy footprint.
• Future Impact: This work validates the potential of dendrocentric computing for low-power, real-time spatiotemporal processing in edge devices, offering a viable alternative to the energy-hungry recurrence and delay mechanisms currently dominating neuromorphic research.

In summary, DendroNN offers a novel paradigm where computation is driven by the detection of specific spatiotemporal patterns rather than global state updates, enabling a new class of highly efficient, event-based AI systems.