Inference with Artificial Neural Networks on Analog Neuromorphic Hardware Johannes Weis (B ) , Philipp Spilger, Sebastian Billaudelle, Yannik Stradmann, Arne Emmel, Eric M¨ uller, Oliver Breitwieser, Andreas Gr¨ ubl, Joscha Ilmberger, Vitali Karasenko, Mitja Kleider, Christian Mauch, Korbinian Schreiber, and Johannes Schemmel Kirchhoff-Institute for Physics, Ruprecht-Karls-Universit¨at Heidelberg, Heidelberg, Germany johannes.weis@kip.uni-heidelberg.de Abstract. The neuromorphic BrainScaleS-2 ASIC comprises mixed-sig- nal neurons and synapse circuits as well as two versatile digital micro- processors. Primarily designed to emulate spiking neural networks, the system can also operate in a vector-matrix multiplication and accumu- lation mode for artificial neural networks. Analog multiplication is car- ried out in the synapse circuits, while the results are accumulated on the neurons’ membrane capacitors. Designed as an analog, in-memory computing device, it promises high energy efficiency. Fixed-pattern noise and trial-to-trial variations, however, require the implemented networks to cope with a certain level of perturbations. Further limitations are imposed by the digital resolution of the input values (5 bit), matrix weights (6 bit) and resulting neuron activations (8 bit). In this paper, we discuss BrainScaleS-2 as an analog inference accelerator and present cal- ibration as well as optimization strategies, highlighting the advantages of training with hardware in the loop. Among other benchmarks, we classify the MNIST handwritten digits dataset using a two-dimensional convolution and two dense layers. We reach 98.0% test accuracy, closely matching the performance of the same network evaluated in software. Keywords: Analog accelerator · Neural network processor · Neuromorphic hardware · Convolutional neural networks · Machine learning · In-memory computing · MNIST 1 Introduction Artificial neural networks (ANN) find application in a wide variety of fields and problems. With networks growing in depth and complexity, the increase of com- putational cost becomes more and more significant [1]. In fact, execution time and power consumption often represent the crucial limiting factors in further scaling and in the application of ANNs [2]. c  Springer Nature Switzerland AG 2020 J. Gama et al. (Eds.): ITEM 2020/IoT Streams 2020, CCIS 1325, pp. 201–212, 2020. https://doi.org/10.1007/978-3-030-66770-2_15