Andreas Toftegaard Kristensen

Andreas Toftegaard Kristensen, Robert Giterman, Alexios Balatsoukas-Stimming et al.

Neural networks have become indispensable for a wide range of applications, but they suffer from high computational- and memory-requirements, requiring optimizations from the algorithmic description of the network to the hardware implementation. Moreover, the high rate of innovation in machine learning makes it important that hardware implementations provide a high level of programmability to support current and future requirements of neural networks. In this work, we present a flexible hardware accelerator for neural networks, called Lupulus, supporting various methods for scheduling and mapping of operations onto the accelerator. Lupulus was implemented in a 28nm FD-SOI technology and demonstrates a peak performance of 380 GOPS/GHz with latencies of 21.4ms and 183.6ms for the convolutional layers of AlexNet and VGG-16, respectively.

Andreas Toftegaard Kristensen, Andreas Burg, Alexios Balatsoukas-Stimming

In this work, we use deep unfolding to view cascaded non-linear RF systems as model-based neural networks. This view enables the direct use of a wide range of neural network tools and optimizers to efficiently identify such cascaded models. We demonstrate the effectiveness of this approach through the example of digital self-interference cancellation in full-duplex communications where an IQ imbalance model and a non-linear PA model are cascaded in series. For a self-interference cancellation performance of approximately 44.5 dB, the number of model parameters can be reduced by 74% and the number of operations per sample can be reduced by 79% compared to an expanded linear-in-parameters polynomial model.