Keywords

Artificial neural networks, Optimization, Printed electronics, Design automation.

Background

Printed Electronics Due to the stretchability, non-toxicity, and ultra-low-cost, printed electronics (PE) dominates the innovative areas such as wearable technology and Internet of Things (IoT) infrastructures. In contrast to silicon-based technology, PE is additive manufacturing that allows high-degree customization and low-amount production, and thus has received considerable interest in edge-AI community.

Artificial neural networks (ANNs) ANNs simulate the human nervous system and have a solid ability to deal with non-linear problems. Besides, the learning-based optimization allows them an extremely high efficiency in designing their parameters. Further, the ultra-simplicity of their computing units (weighted-sum & activation) are ideal for the hardware implementation in PE. A promising approach it is the printed neuromorphic computing system.

Printed neuromorphics Despite the availability of a large number of computing units in the field of ubiquitous computing and edge-AI (such as Arduino and mobile phones), the cost and extravagant computing resources can still be saved by printed neuromorphic systems. Printed neuromorphics are the hardware implementation of ANN using resistors, transistors, etc. However, the emerging printed neuromorphics are not perfect. Material imperfections and limited processing technology render well-designed printed neuromorphics difficult. As printed neuromorphics are based on analog circuits, fabrication errors in the resistors can bias the calculation of the designed output, and aging problems of the resistors can cause the output to vary during usage.

Goal of thesis The outputs of the printed neuromorphics depend only on the current input and cannot register or analyze the previous inputs, e.g., the printed neuromorphics cannot even tell whether the input is increasing or decreasing over time. This weakness makes the utilization of printed neuromorphics significantly limited. We, therefore, came up with the idea of adding multiple filters (band pass, low pass, high pass, etc) before the input to the printed neuromorphics. Due to the equivalence of the frequency- and time-domains, the filtered signals then take the temporal information, which enables the printed neuromorphics to deal with time series.

Your jobs

• Research state of the art printed electronic technology

• Research the methods for signal processing

• Implement a training framework for printed neuromorphics

• Design filters for printed neuromorphics

• Implement a training framework for printed neuromorphics and filters

• Evaluate the proposed framework

• Write thesis

We provide you

• Chance & experience in field of artificial neural networks, design automation, and optimization.

• Intensive and interdisciplinary support

• A pleasant working atmosphere and constructive cooperation

• Possibility to publish this work as one of the co-authors

We expected

• Students major in Electrical Engineering, Mechatronics, Informatics, and other related majors

• Independent thinking and working

• Strong knowledge of Python (specifically Pytorch)

• Strong knowledge in artificial neural networks

• Basic knowledge in signal-processing

• Basic knowledge in electronics

Contact

If you are interested in this work, please contact Haibin Zhao (haibin.zhao@kit.edu).

Reference

Weller D D, Hefenbrock M, Beigl M, et al. Realization and training of an inverter-based printed neuromorphic computing system[J]. Scientific reports, 2021, 11(1): 1-13.

Folland, G. B. Fourier Analysis and Its Applications. Pacific Grove, Calif: Wadsworth & Brooks/Cole Advanced Books & Software, 1992.