💻 computer science

Unified Complex-valued Neural Network: A Magnitude-Phase Computational Model for Event-Driven Neuromorphic Learning

This paper introduces the Unified Complex-valued Neural Network (UCNN), a novel model that integrates continuous magnitude encoding with phase-driven event generation via an asymmetric complex-valued state, enabling efficient, sparse, and interpretable spatiotemporal learning through both unified backpropagation and an event-driven adaptive learning rule for neuromorphic applications.

Original authors: Reza Ahmadvand, Sarah Safura Sharif, Yaser Mike Banad

Published 2026-06-30

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Reza Ahmadvand, Sarah Safura Sharif, Yaser Mike Banad

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to teach a robot to follow a moving car. You have two very different ways to give the robot instructions, but both have a major flaw.

The Problem: The "Photo" vs. The "Flash"

The "Photo" Approach (Traditional AI): Imagine a robot that takes a high-resolution photo of the car every single second. It sees everything clearly and knows exactly where the car is. However, it's slow and wasteful. It processes the whole picture even when the car hasn't moved an inch. It's like taking a photo of a still pond every second just to make sure the water hasn't changed.
The "Flash" Approach (Spiking Neural Networks): Now imagine a robot that only "wakes up" when it sees something move. It's incredibly fast and energy-efficient because it ignores everything static. But, it only sees "blips" of motion. It knows when something moved, but it's bad at knowing how far or how strong the movement was. It's like a security camera that only records a split-second flash when a shadow passes, making it hard to tell if a person walked by or a bird flew by.

For a long time, engineers had to build two separate systems to handle these two types of information, or they had to choose between being accurate (but slow) or being fast (but imprecise).

The Solution: The "Pulsing Heart" (The UCNN)

This paper introduces a new kind of "brain cell" called the Unified Complex-valued Neuron (UCN). Think of this neuron not as a simple switch, but as a pulsing heart.

A heart has two distinct but connected parts:

The Beat (Magnitude): This is the strength of the pulse. In the new model, this represents the value or intensity of the signal (like "how bright" the light is or "how far" the object moved). It's continuous and smooth, just like a traditional photo.
The Rhythm (Phase): This is the timing of the beat. In the new model, this represents when an event happens. The heart beats faster or slower depending on what it senses. When the rhythm hits a certain point (a full circle), it fires a "spike" or an event.

How It Works Together

Instead of having two separate systems, this new network (called UCNN) combines them into one.

The Magnitude (strength) tells the network what is happening.
The Phase (timing) tells the network when to react.

It's like a drummer who can play a steady beat (timing) while also changing how hard they hit the drum (strength). The drummer doesn't need two different people to do this; it's one person doing two things at once.

The "Smart" Learning Trick

Teaching a system that changes its own rhythm over time is usually very hard and computationally expensive (like trying to rewind a movie frame-by-frame to fix a mistake).

The authors created a new learning rule called EAPL (Event-Driven Adaptive Phase Learning).

The Analogy: Imagine you are trying to fix a mistake in a song. Instead of listening to the whole song again and again to find the wrong note, you have a "magic ear" that instantly tells you exactly which note was off and how to fix it, skipping all the parts that were already perfect.
The Result: This makes the training process much faster and more efficient, allowing the network to learn complex patterns without getting bogged down by heavy calculations.

Did It Work?

The researchers tested this new "pulsing heart" brain on two specific challenges:

Chasing a Moving Object: They asked the network to track a ball moving in a circle.
- The old "Flash" system (SNN) was jittery and missed the exact spot.
- The new "Pulsing Heart" system (UCNN) tracked the ball smoothly and accurately, almost as well as the slow "Photo" system, but without the heavy processing cost.
Predicting Chaos (The Lorenz Attractor): They asked the network to predict the movement of a chaotic, swirling system (like a weather pattern).
- The "Photo" system was okay, but not perfect.
- The "Flash" system struggled to keep up with the complex swirls.
- The "Pulsing Heart" system (UCNN) was the best at predicting the chaotic path, staying stable and accurate.

The Bottom Line

This paper presents a new way to build AI that gets the best of both worlds: the precision of traditional computers and the speed/efficiency of biological brains. By using a single neuron that understands both "strength" and "timing" simultaneously, the UCNN can learn complex, moving tasks accurately without needing massive amounts of computing power. It's a step toward making AI that is as smart as it is efficient.

Technical Summary: Unified Complex-valued Neural Network (UCNN)

Problem Statement
The paper addresses a fundamental tension in neural computing paradigms: Artificial Neural Networks (ANNs) offer high-precision continuous-valued representations but lack intrinsic event-driven temporal processing, while Spiking Neural Networks (SNNs) provide efficient, asynchronous event-driven computation but often sacrifice graded signal representation through binary spike encoding. Existing hybrid architectures attempt to bridge this gap but frequently rely on separate modules, parallelized networks, or complex synchronization mechanisms, which introduce design overhead and limit the tight coupling of static context and dynamic temporal information. Furthermore, pure SNNs struggle with stable learning for robust dynamical-system tasks due to the reduction of information to binary events.

Methodology
The authors propose the Unified Complex-valued Neural Network (UCNN), built upon a novel neuron model called the Unified Complex-valued Neuron (UCN).

Neuron Architecture: The UCN utilizes an asymmetric complex-valued internal state, $z(t) = r(t)e^{i\theta(t)}$ .
- Magnitude ( $r$ ): Encodes continuous signal strength using a rectified hyperbolic tangent function of the pre-activation. This channel provides bounded, continuous activation similar to ANNs.
- Phase ( $\theta$ ): Acts as an internal timing variable governed by leaky, driven first-order dynamics (analogous to a Leaky Integrate-and-Fire neuron). The phase evolves continuously and gates the emission of events. When the phase crosses a threshold (e.g., $2\pi$ ), a spike is generated.
- Output: The neuron emits a "valued event" $y(t) = s(t) \cdot r(t)$ , where $s(t)$ is the phase-gated spike indicator. This decouples event timing (governed by phase) from signal amplitude (governed by magnitude).
- Reset Mechanisms: The model supports hard reset (to zero), modulo reset (subtracting threshold), and soft reset (decay) strategies to manage phase continuity and firing regularity.
Training Frameworks:
1. Unified BP-BPTT: A foundational training rule combining Backpropagation (BP) for the continuous magnitude pathway and Backpropagation Through Time (BPTT) with surrogate gradients for the event-governed phase pathway. This allows end-to-end optimization of both "what" (magnitude) and "when" (phase) information.
2. Event-Driven Adaptive Phase Learning (EAPL): To reduce the computational cost of explicit temporal unrolling required by standard BPTT, the authors introduce EAPL. This method reformulates learning as a backward-time dynamical process using an adjoint state. It propagates task-relevant error sensitivity through the phase dynamics without explicitly unrolling the entire computational graph, significantly reducing backward-pass operations while maintaining compatibility with the UCNN architecture.

Key Contributions

Novel Architecture: Introduction of the UCNN, a magnitude-phase event-valued model that unifies continuous state accuracy with sparse, phase-governed event timing in a single computational structure, eliminating the need for parallelized hybrid modules.
Unified Learning Rule: Development of a BP-BPTT framework that jointly optimizes magnitude and phase pathways, treating them as a unified computational graph.
Efficient Learning Algorithm: Proposal of the EAPL rule, which utilizes adjoint-style recursion to propagate phase sensitivity, reducing computational burden compared to explicit unrolling.
Validation: Demonstration of the model's efficacy on two distinct tasks: mixed-modality object tracking and the prediction of chaotic nonlinear dynamics (Lorenz attractor).

Experimental Results
The paper evaluates UCNN against standard ANNs and SNNs on two tasks:

Object Tracking (Moving Object on Circular Path):
- Performance: The UCNN achieved tracking accuracy between that of the ANN (highest accuracy, dense input) and the SNN (lowest accuracy, binary input). Specifically, UCNN significantly reduced the Root Mean Square Error (RMSE) and maximum tracking error compared to the SNN, while maintaining a more stable error profile with less jitter.
- Activity: UCNN exhibited intermediate sparsity—sparser than the dense ANN but more active than the purely spiking SNN.
- Trajectory: The UCNN trajectory closely followed the ground truth, outperforming the SNN in capturing local deviations and maintaining consistency.
Lorenz Attractor Prediction:
- Performance: In learning the coupled nonlinear dynamics of the Lorenz system, the UCNN trained with EAPL achieved the lowest training loss and the smallest RMSE across all state components ( $x, y, z$ ) compared to both ANN and SNN baselines.
- Stability: The UCNN demonstrated superior stability, with the narrowest error distribution and lowest maximum error. It successfully reconstructed the attractor geometry with minimal deviation, whereas the SNN showed significant distortion and the ANN showed moderate deviations.
- Activity: Similar to the tracking task, UCNN maintained event-driven sparsity while preserving richer information flow than the SNN.

Significance and Claims
The paper claims that the UCNN successfully bridges the gap between continuous-valued precision and event-driven temporal processing. By integrating magnitude and phase into a single asymmetric complex-valued state, the model eliminates the architectural overhead of traditional hybrid networks. The authors assert that the UCNN provides a robust, scalable framework for neuromorphic and edge-AI applications, offering accurate spatiotemporal learning while preserving the energy efficiency and sparsity of event-driven computation. The results suggest that the magnitude-phase formulation is effective not only for visual perception tasks but also for learning complex, chaotic dynamical systems, positioning the UCNN as a general neural computational model.

More like this