Abstract
DOAJ is a unique and extensive index of diverse open access journals from around the world, driven by a growing community, committed to ensuring quality content is freely available online for everyone.
Introduction
While the acronym "CNN" is now synonymous with Convolutional Neural Networks in the context of deep learning, its origins in signal processing refer to Cellular Neural Networks. This editorial introduces a special issue of the EURASIP Journal on Advances in Signal Processing dedicated to this original "Cellular" paradigm. CNNs provide a framework for massively parallel, non-linear spatiotemporal dynamics, a field often referred to as Cellular Wave Computing.
Addressing the Von Neumann Bottleneck
Traditional von Neumann architectures, characterized by serial processing, are often inefficient for complex, real-time spatiotemporal tasks. Simulating wave propagation, modeling biological vision, or solving NP-hard optimization problems requires a computing paradigm that handles non-linear dynamics natively. CNNs, defined by Professor Leon Chua in 1988, consist of regularly spaced cells with local connectivity. This architecture is exceptionally well-suited for high-speed parallel processing and physical hardware implementation, bridging the gap between abstract mathematical models and real-world engineering applications.
Methodologies in Cellular Wave Computing
The research presented in this issue synthesizes several key approaches to CNN implementation:
- Mathematical Modeling: Solving Partial Differential Equations (PDEs) and analyzing the stability and convergence of non-linear dynamics.
- Digital Emulation: Implementing CNN dynamics on FPGAs, multicore microprocessors, and GPUs to leverage modern digital performance and Moore’s Law.
- Analog/Mixed-Signal Design: Exploring CMOS-3D technology and nanodevices to create custom hardware that mimics biological neural circuits, such as retina models.
- Software Arrays: Developing fine-grained parallelism on standard microprocessors through software-based processor arrays.
Key Insights and Application Domains
The field has transitioned from static image processing to modeling complex spatiotemporal phenomena, such as higher brain functions and biological motion. A notable trend is the shift toward digital platforms; five of the nine featured papers utilize FPGAs or GPUs, indicating that digital performance has reached a level where it can effectively emulate specialized CNN ASICs.
Application domains highlighted include:
- Human EEG Analysis: Using CNNs to study synchronization phenomena in brain activity.
- Supersonic Flow Simulations: Testing FPGA-based CNN realizations with two-dimensional data.
- Biological Vision: Developing retina models for motion estimation and visual attention.
- Optimization: Demonstrating that CNN-based computers can outperform conventional digital systems in specific NP-hard problems.
Results and Hardware Milestones
The editorial highlights several quantitative and qualitative milestones achieved through these methodologies. CNNs have proven effective for real-time stability and convergence in signal processing tasks. Specific hardware successes include the implementation of retina models and supersonic flow simulations on FPGA platforms. Furthermore, the integration of GPU acceleration has provided a vital bridge, allowing researchers to utilize high-level programming while maintaining the benefits of massively parallel execution. The realization of analog processor arrays in conventional CMOS technology specifically for motion estimation marks a significant step in biological emulation.
Future Outlook
Despite the accessibility of digital platforms, the authors argue that custom analog or mixed-signal hardware remains essential for the most demanding applications, such as high-speed visual tracking. The next major leap is expected to come from CMOS-3D technology and nanodevices, which will push the boundaries of visual attention systems by overcoming the physical constraints of 2D silicon layouts. There remains a critical need for research that occupies the middle ground between pure mathematical theory and direct hardware application to fully realize the potential of Cellular Wave Computing.
Summary updated