FPGA Implementation and Analysis of Non-Linear Activation Functions for Implicit Neural Representations

Implicit Neural Representations (INR) have found applications in various sectors like health care (computed tomography) and autonomous navigation (image processing). While there are many use cases of INRs, the non-linear activation functions are a key differentiating part in such INRs. Hence, this paper provides novel FPGA based implementations of the non-linear activation functions, namely Sine, Gaussian, and Complex Gabor Wavelet and their derivatives. All three non-linear activation functions and their derivatives were designed using AMD-Xilinx Vivado HLS 2019.2 and were synthesized, run implementation, and generate bit-stream using AMD-Xilinx Vivado 2019.2. Xilinx PYNQ framework was used for validation on the target FPGA embedded platform (Avnet Ultra96 v2 - Xilinx ZU3EG Zynq Ultrascale + MPSoC). Further, a detailed comparison highlighting their importance with respect to some key parameters like silicon area, power and timing has been provided to allow design engineers and researchers working in this field to make an appropriate choice based on the application considered.