A Design Flow Framework for Fully-Connected Neural Networks Rapid Prototyping.

The current work deploys a framework for rapid prototyping of Fully-Connected Neural Networks (FCNs). The scope is to provide an automatic design flow that generates a template-based VHDL code considering the accuracy, the resource utilization and the design complexity. More precisely, the deployed tool incorporates hardware optimizations in the implementation of the multiplications, the activation function and the definition of the fixed-point types providing user-defined configurations thought a GUI. The FCNs of two applications (Alexnet and Lenet) were implemented to evaluate our approach. The results seem promising and prove the design flexibility of our framework generating optimized code that exceeds the 10K lines for each hardware instance within a few hours, while preserving low levels of latency that does not exceed 400 cycles for our applications.