Neuromorphic Memristive VLSI Architectures for Cognition (NMVAC)

• Applicant: Professor Dr. Martin Ziegler, since 4/2020
• Subject Area: Electronic Semiconductors, Components and Circuits, Integrated Systems, Sensor Technology, Theoretical Electrical Engineering
• Term: from 2020 to 2024
• Project identifier: Deutsche Forschungsgemeinschaft (DFG) - Project number 432009531

The realization of biological inspired spiking neural networks is strongly hindered by the exponentially increasing gap between processor speed and memory bandwidth characteristic of conventional von Neumann computing architectures. In these networks processing elements, namely the neurons, are interconnected with so called synapses. The simulation of both elements requires an intense memory use for the access and update of state variables. Therefore, real-time simulation of large scale networks is constrained by the von Neumann bottleneck. In contrast, the highly parallel and energy efficient computing substrate offered by neuromorphic analog Very Large Scale Integration (VLSI) systems can scale up without affecting simulation time, thanks to the co-location of memory and computation.This project will contribute to the recent cutting-edge research efforts for the realization of fully integrated CMOS/memristive systems optimized for the emulation of learning neural networks. We will use the stochastic switching of binary RRAM devices to implement learning algorithms inspired by the nervous system. The devices will be embedded in state-of-the-art subthreshold analog circuits on CMOS chips. This combination will lead to the construction of learning systems with unprecedented low power consumption, always operating in real time and suitable for on-line learning.In this proposal, we envision a well defined path towards the achievement of our research goals. First, we will extensively characterize the HfOx-based memristive devices. Based on this characterization, we will develop an accurate mathematical model of the stochastic switching behavior. This model will be instrumental for the realization of faithful SPICE simulation of hybrid CMOS/memristive circuits as well as the implementation of realistic spiking neural network simulations supporting the hardware design. The core activities of the project will include the design and fabrication of two hybrid memristive/CMOS chips. The final system will support full flexibility in terms of network architectures, therefore supporting the implementation of both feed-forward and recurrent structures. Its learning and cognitive performances will be thoroughly evaluated in the context of classification, pattern recognition, pattern completion and pattern separation tasks. Results gathered and technology developed within this project will have a major impact in the neuromorphic engineering field promoting innovative uses of RRAM devices and advancing the state-of-the-art of subthreshold analog circuits for spiking neural networks. On a broader scale, this project has high potential for contributing to the emerging field of neuromorphic intelligence and its killer applications, comprising the use of dedicated cognitive systems for edge-computing.

• DFG Programme: Research Grants
• Ehemalige Antragstellerin: Professorin Elisabetta Chicca, Ph.D., until 4/2020