An efficient and expandable hardware implementation of multilayer cellular neural networks

This paper proposes a new CNN architecture conceived for hardware implementation of complex ML-CNNs on programmable devices. The architecture is completely modular and expandable, and includes advanced features such as non-linear templates, time-variant coefficients or multi-layer structure. We also present an implementation platform based on the pre-designed but user-configurable FPGA processing modules that inherit the modularity and expandability of the logical architecture. All the modules share the same, properly designed, I/O interface, so the platform can be configured to accommodate CNNs of any size or structure, composed of a number of processing blocks that can be physically distributed over several FPGA boards. Our Carthagonova architecture makes use of a temporal processing approach with a super-pipelined unfolded cell structure, leading to the maximum degree of parallelism while still keeping the most efficient use of FPGA resources. Both the CNN architecture and the hardware platform have been validated by the implementation of a real-time video processing system, showing that they conform a valuable set of tools for the development of CNN-based applications.