TU Dresden
A joint project with Uni Leipzig DE EN

MemAI: Memristive Dynamic Models for Smart AI Systems

Mentored by Ronald Tetzlaff, Thomas Mikolajick
at Chair of Fundamentals of Electrical Engineering, Institute of Circuits and Systems, TU Dresden

Memristive technologies are expected to enable various novel energy-efficient applications for neuromorphic computing to compete with the huge amounts of data processed in neural networks with a new in-memory computing topology that overcomes the so-called memory wall between processing and memory chip. Various memristive devices have successfully demonstrated that they can perform highly parallel matrix-vector multiplication (MVM) and logic algorithms in crossbar arrays.

However, compact modeling and efficient read/write circuits are required to simulate state-of-the-art neural networks and predict working applications with current memristor materials, while overcoming real device challenges such as endurance, storage, and variability on the one hand, and enabling fully analog memristive systems including neuron activation on the other. Proper modeling of devices and systems facilitates analysis of dynamic layers, such as the LSTM layer often used in Recurrent Neural Networks (RNNs) and Transformers that consider tunable synaptic weights in memristive crossbars with nodes connected to themselves via feedback loops.

The PhD topic investigates the modeling and simulation of dynamic memristive circuits enabling modern dynamic neural networks such as RNNs. The research challenges that are crucial in this context are for example:

  • How can memristive systems be used to implement modern neural network structures such as RNNs and Transformers that require regular updating of synaptic weights with feedback loops?
  • How can inappropriate properties of certain devices, such as low endurance, short retention time, and high variability, be incorporated and minimized in the design of advanced memristor AI systems
  • What are the most promising circuits for implementing fully-analog memristive neural networks including neuron activation and overcoming ADC/DAC conversion.
  • How can memristor-based in-memory computing be employed for performing both MVM and logical operations and serving as versatile accelerator in AI systems?

Work environment

You will be working in the research group of Prof. Dr. Ronald Tetzlaff at TU Dresden providing a versatile high performance computing lab for running memristive system simulations. The Chair of Fundamental of Electrical Engineering has an inspiring atmosphere with experienced and international PhD and PostDoc researchers and holds active connections to the world-leading researchers in the field of memristors, e.g. Leon Chua (UC Berkeley) and R. Stanley Williams (Texas A & M University).

You will be supported by Prof. Dr. Thomas Mikolajick enabling access to experimental data on real memristor devices for model verification. The chair of Nanoelectronics at TU Dresden runs a fully equipped state of the art research cleanroom for the fabrication of test structures and test devices.

Prerequisites

To conduct this research, you should hold a very good university degree (MSc or an equivalent) in electrical engineering or related disciplines (such as computer science, nanoelectronics, mechatronics). Excellent skills and practical experience in one or more of the research areas, memristive device modeling and machine learning, is beneficial. You are able to collaborate well in an interdisciplinary and international environment.