An Overview of Arithmetic Adaptations for Inference of Convolutional Neural Networks on Re-configurable Hardware

Convolutional Neural Networks (CNNs) have gained high popularity as a tool for computer vision tasks and for that reason are used in various applications. There are many different concepts, like single shot detectors, that have been published for detecting objects in images or video streams. However, CNNs suffer from disadvantages regarding the deployment on embedded platforms such as re-configurable hardware like Field Programmable Gate Arrays (FPGAs). Due to the high computational intensity, memory requirements and arithmetic conditions, a variety of strategies for running CNNs on FPGAs have been developed. The following methods showcase our best practice approaches for a TinyYOLOv3 detector network on a XILINX Artix-7 FPGA using techniques like fusion of batch normalization, filter pruning and post training network quantization.

Polarization-Dependent Loss of Optical Connectors Measured with High Accuracy (<0.004 dB) after Cancelation of Polarimetric Errors

State-of-the-art polarimeter calibration is reviewed. Producing many quasi-random polarization states and moving/bending a fiber without changing power allows finding a polarimeter calibration where the degree-of-polarization reaches unity and parasitic polarization-dependent loss is small. Using a polarization scrambler/transformer and a polarimeter a device-under-test can be characterized. Its Mueller matrix can be decomposed into a product of a nondepolarizing Mueller-Jones matrix times a purely depolarizing Mueller matrix. Test polarizations may drift over time. With help of an optical switch the reference device can be measured against an internal reference path. Later, with possibly different test polarizations, the actual device-under-test is measured against the internal reference. Polarization drift and need for repeated reference device measurement are thus overcome. When a patchcord is inserted, connector PDL can be measured, provided that errors are calibrated away, again by fiber moving/bending. Experimentally we have measured PDL with errors <0.004 dB. This easily suffices to measure connector PDL, which is demonstrated. PDL >60 dB was measured when the device under test was a good polarizer. A 20 Mrad/s polarization scrambler with LiNbO3 device generates the test polarizations. The polarimeter can sample at 100 MHz and can store 64M Stokes vectors. During laser frequency scans Mueller matrices can be measured in time intervals as short as 5 us.