Schematic Generation of Programmable Analog Neural Networks for Signal Proccessing

Konferenz: SMACD / PRIME 2021 - International Conference on SMACD and 16th Conference on PRIME
19.07.2021 - 22.07.2021 in online

Tagungsband: SMACD / PRIME 2021

Seiten: 4Sprache: EnglischTyp: PDF

Aul, Florian; Krischker, Lukas; Schmalhofer, Sascha; Hedrich, Lars (Institute for Computer Science, Goethe University Frankfurt, Germany)
Katsaouni, Nikoletta; Schulz, Marcel H. (Institute for Cardiovascular Regeneration, Goethe University Frankfurt, Germany)

This paper presents a methodology for generating analog neural networks (ANNs) from trained TensorFlow models using an intermediate description of the network. All needed circuitry and helper circuits are available in a block library provided for the operation in an energy efficient low power region. The weights of the ANN are programmable by the corresponding bitstream, which is also generated by the proposed approach. The feasibility of the approach is demonstrated with two examples, the larger being an ANN for arrhythmia detection of electrocardiograms (ECGs) with 2,570 neurons and 10,042 weights.