15:00-15:45

The fourth element (or the rediscovered memristor)

Ομιλητής: Valeri Mladenov , Professor at the Technical University of Sofia, Bulgaria

Abstract: Memristor is the fourth fundamental circuit element which was predicted in the 1970s and rediscovered in HP labs in 2008. The HP memristor, is constructed using a nanometer scale TiO2 thin-film that comprises a doped and an undoped region. It is a novel circuit element with very good memory and switching properties, nano-size dimensions, low power consumption, and good compatibility with complementary metal oxide semiconductor (CMOS) integrated circuits. The potential applications of memristors are not only limited to artificial biological systems and nonvolatile RAM, but also extend to reconfigurable nanoelectronics. Furthermore, memristors introduce novel paradigms in the application-specific integrated circuits (ICs) and field-programmable gate arrays (FPGAs) domains. The implementation of memristors in ICs offers unprecedented memory capacity and device density while occupying a significantly smaller area. Memristors could be used for energy-efficient storing and processing information in high-density integrated chips applicable in nonvolatile memories, artificial neural networks, and neuromorphic computing circuits. In this lecture, an introduction to memristors is given, along with the presentation of mathematical and SPICE models for these devices. Additionally, an overview of the potential applications of memristors is provided.

15:45-16:00

Διάλειμμα

16:00-16:30

Memristor-Based Classical Neural Network Implementations with Adjustable Synaptic Weights

Ομιλητής: Georgi Tsenov, Associate Professor at the Technical University of Sofia, Bulgaria

Abstract: The memristors are innovative electronic elements with nano-sized structure and with very good memory and switching abilities. They have very low power consumption and a good compatibility to CMOS integrated chips, and they could be used in neural networks, memories, and many other schematics. Most of the memristor implementations are with crossbar configuration, suitable for CNNs, but recently chips with 8, 16 or 32 individually accessible memristors was released by Knowm, providing the option to build classical neural network designs. A classical artificial neural network can be implemented as hardware analog device with memristor-based synapses that can be tuned to certain scaled resistance levels, opening the option for analog computing and neural network coefficient reprogramming. In this network, each synapse is realized with only one memristor, thus providing a higher reduction in circuit complexity and with main benefit of that individual memristor resistance value can be adjusted with external control voltage signals. The summing and scaling component implementations are based on op-amps and memristors. We use the most common hardlim or logarithmic-sigmoidal activation function and it is realized by a voltage-controlled source. The operation of the proposed memristor neural network is analyzed and simulated in both LTSPICE and MATLAB, and the derived results are compared and verified successfully with hardware implementations on breadboards. The proposed memristor-based neural network is a significant step for engineering low power complex networks in very high-density integrated circuits and chips.

16:30-17:00

Automatic SPICE memristor model estimation based on measurement data from existing memristors

Ομιλητής: Georgi Tsenov, Associate Professor at the Technical University of Sofia, Bulgaria

Abstract: The real-world production memristor devices produce a modelling challenge due to having a memristor element variance and henceforth a measured difference between breadboard made circuits and LTSPICE or MATLAB simulated circuits with classical models, resulting in a variance owed to the difference in existing memristor models and the real world hardware. When a new memristor device is studied usually no model is provided and it is useful if there is a tool that automatically can update existing models with the production memristor parameters. This motivated us to implement a procedure realized in MATLAB that takes real world measured data, approximates the memristor parameters and provides a MATLAB and LTSPICE models that are precise representation of the real world memristor device.