Quaternion-Valued Neural Network Dynamics

Quaternion-valued neural networks employ quaternions – four-dimensional hypercomplex numbers – as fundamental representational units, allowing weights, inputs and activations to be encoded compactly while capturing interlinked correlations among multidimensional features. The core mathematical operation is the Hamilton product, a non-commutative multiplication that entwines real and imaginary components to preserve geometric and spatial relationships. This framework offers several advantages over traditional real-valued networks: a dramatic reduction in free parameters, inherent coupling of vector channels, and the ability to process vectorial data (such as colour images, three-axis sensor readings or spatial orientations) in a single algebraic domain. Dynamical analyses of quaternion networks address stability, synchronisation and response to delays or stochastic disturbances. Techniques range from Lyapunov-based methods for global exponential stability to decomposition into equivalent real-valued systems for tractable analysis of convergence and robustness. Quaternion dynamics also extend to memristive and Hopfield architectures, yielding enhanced capabilities for secure communication, associative memory and time-series prediction. As research bridges theoretical foundations with practical applications, quaternion-valued dynamics are emerging as a versatile tool for efficient representation and processing in high-dimensional signal and data-modelling tasks.

Research from Nature Portfolio

No recent Nature Portfolio content available.

Research from all publishers

Several recent contributions have advanced learning and architectural components for quaternion networks. A lightweight family of parameterized hypercomplex convolutional models demonstrates that hypercomplex layers – including quaternion specialisations – can learn convolutional rules directly from data without a predefined algebraic domain. These networks operate with fewer parameters than comparable real or quaternion networks, achieving competitive performance on both image and audio benchmarks. In parallel, a comparative study of quaternion backpropagation algorithms has unified disparate optimisation approaches by evaluating four distinct gradient-calculus frameworks, including novel generalised HR-calculus. Rigorous experimental comparison across regression tasks reveals that specific quaternion-tailored learning rules can offer significant gains in accuracy without imposing additional computational cost. Finally, novel quaternionic activation functions have been proposed to respect the full algebraic structure rather than applying split real functions to individual components. By modulating quaternion magnitude or phase jointly, these activations improve gradient flow and leverage the Hamilton product’s coupling, yielding consistent performance improvements on image classification challenges over conventional split-ReLU and split-Tanh schemes.

Quaternion-Valued Neural Network Dynamics publication trend

The graph below shows the total number of articles in quaternion-valued neural network dynamics across all publications each year (not limited to Nature Index journals).

Technical terms

Quaternion: Four-dimensional hypercomplex number system comprising one real and three imaginary units, used to encode input, weight and activation in a single algebraic structure.

Hamilton product: Non-commutative multiplication operation for quaternions, enabling coupling between quaternion components and preserving geometric relations in vector-valued data.

Hypercomplex convolution: Generalisation of convolutional operators to algebraic domains beyond real numbers, permitting multidimensional filter interactions conforming to quaternionic algebra.

Split activation: Strategy of applying real-valued activation functions separately to each quaternion component, often contrasted with quaternionic activation functions that respect the full algebra.

References

1. PHNNs: Lightweight Neural Networks via Parameterized Hypercomplex Convolutions. IEEE Transactions on Neural Networks and Learning Systems (2024).
2. A comparison of quaternion neural network backpropagation algorithms. Expert Systems with Applications (2023).
3. Improving quaternion neural networks with quaternionic activation functions. Knowledge-Based Systems (2024).
4. Global exponential synchronization of high-order quaternion Hopfield neural networks with unbounded distributed delays and time-varying discrete delays. Mathematics and Computers in Simulation (2022).
5. Stochastic Memristive Quaternion-Valued Neural Networks with Time Delays: An Analysis on Mean Square Exponential Input-to-State Stability. Mathematics (2020).