Annotation of the results obtained in 2018
During the reporting period, a cycle of research activities has been completed that ensures the development and implementation of a neural interface based on the integration of a memristor artificial neural network (ANN) with a living neural network in a culture of dissociated hippocampal brain cells. The following tasks have been reached in accordance with the research plan: development of the ANN prototype software based on memristor matrices in order to use synaptic plasticity as part of the learning system; study of the possibility of functioning of memristors within the framework of ANN models based on memristor matrices and external electronic nodes, including the interface with a multielectrode system interfaced with a neural network of cultured brain cells; studying the possibility of building a feedback “memristor – neural network of culture of dissociated brain cells”. Monitoring and generalization of scientific and technical information on the subject of the project has led to the conclusion that, despite the steady trend in the transition from traditional ANNs based on programmable memristive weights to spiking neural networks with self-organization of memristive connections, both directions are being actively developed. The analysis of foremost reports confirms the full compliance of the tasks and results of the project with the world level, as well as the validity of the previously made conclusion about the need to maintain a balance in the combination of different approaches. The first one and the main one is to demonstrate the capabilities of a one-board-integrated double-layer perceptron based on programmable memristive devices as a classifier of informative characteristics of bioelectric activity, which makes it possible to implement a living neural network – memristor ANN adaptive neural interface in the form of a compact autonomous device. The second approach, which is developed in parallel in our project and consists in the development of unconventional neural network architectures based on the stochastic dynamics of memristive devices, creates an important foundation for a transition to a qualitatively new level in the field of creating biosimilar memristor systems, as well as beyond the scope of the current project. In the framework of the first approach, the algorithmic and control software of the ANN prototype was developed based on the matrix of memristive microdevices, which was tested during programming of the resistive states of memristors corresponding to weights determined during the training of the ANN computer model. The processing and classification of electrical activity of biological networks in the memristor-based ANN was tested directly on the example of informative characteristics obtained in special experiments in response to given stimuli in the culture of dissociated hippocampal cells. Within the framework of the second approach, the automated measurements by using the developed hardware and software complex based on National Instruments DAQ data-acquisition system allowed for a more detailed study of the nature of alternative types of synaptic plasticity of memristive devices based on the adaptive response of memristor to spike-like signals and neuronal activity. The modified memristor stochastic response model allowed us to simulate more complex network architectures based on neuron-like generators and memristive devices. The stochastic dynamics of memristive devices is most pronounced when exposed to signals registered directly in the culture of neuronal cells. Statistical analysis has showed that preferred states are highlighted in the histograms of resistive states, which indicate the complex (multistable) energy profile of the memristor as a dynamical system, which also evolves over time. This phenomenon opens up a wide area for further research and requires theoretical description using the rich tools of statistical physics, which is obviously beyond the scope of the current project. The obtained results allowed us to make practical recommendations on the ways of building the feedback “memristor – neural network of culture of dissociated brain cells” allowing to regulate the activity of a biological neural network depending on changes in its response to given stimuli. The structural-functional scheme of the adaptive neurointerface of automatic registration and stimulation of cells with feedback includes the following elements based on various developed and tested software and hardware systems and the technology of neuronal activity delivering: memristor ANN; algorithmic and control software to provide the ANN learning process; interface software; a system for recording and stimulating bioelectric activity; software module for controlling the stimulation of neuronal culture activity. The whole set of elements can be placed in a single package, so the neural interface can be used in compact robotics control systems. The proposed scheme is universal and allow any improvements in individual modules in order to increase their functional characteristics. Thus, during the implementation of the proposed approach to the creation of an adaptive neural interface, the following difficulties have emerged that can be solved in the course of further research and development. First, the capabilities and functionality of traditional neural network architectures based on programmable memristive weights are limited by the size of memristor array, the increase of which is constrained not by low scalability (the minimum size of memristive element may be on the order of a nanometer), but the insufficiently reproducible parameters of devices due to the stochastic nature of the resistive switching phenomenon. The solution to this problem can be based on the engineering of memristive nanostructures and the use of automated programming techniques with active feedback, allowing to form and maintain large arrays of memristive devices. The controllability of the response of living neuronal cultures is a major problem and can be ensured by the spatial separation of neurons in culture using microfluidic technologies that are compatible with the standard microelectrode approaches for recording and stimulating bioelectric activity. Thus, all the tasks set at the third stage of the project have been reached to ensure the achievement of the main goal of the project and allowed to outline further ways for the development of its results. Comparison of the obtained results with the current scientific and technical background allows us to evaluate the level of research performed as corresponding to the best world achievements in the field of creating neuromorphic and neurohybrid systems based on memristive nanomaterials. The important role of the completed project is that it has provided a stable 'springboard' for the deployment of a number of new directions related to the creation of hybrid analog-digital neuroprocessors based on memristors, brain-like architectures based on neuronal synchrony and self-organization, as well as solving fundamental problems of describing dynamical memristive systems using the newest methods of statistical analysis. Information about the project on the Internet is available in the Research Gate profile of principal investigator: https://www.researchgate.net/project/Biosimilar-memristive-systems-coupled-with-neuronal-biocultures.

 

Publications

---

Annotation of the results obtained in 2016
During the reporting period, the cycle of research activities was performed that provides the development and approbation of the main elements of a neurointerface, which couples the artificial neural network (ANN) based on a weight matrix of memristive devices with the living neural network in a culture of dissociated hippocampal cells of the brain. The cycle included the following activities: formation of thin film memristive structures of the ‘metal-oxide-metal’ type, of different structure, topology and geometric parameters; selection on the basis of experimental study and computer simulation of constructive variants of memristive structures for the implementation of synaptic plasticity and their integration into the ANN; development and verification of a neural network computer model based on the artificial (electronic) neurons and a matrix of weights composed of selected variants of memristive structures; selection and testing of the interface systems and approaches for the registration and transfer to ANN of signals that characterize in the best way the response of living cultures of dissociated brain cells to a given stimulus; cultivation of dissociated cells on multi-electrode matrices (MEM) with the registration and stimulation of activity, creation of a database of neural activity. Basing on the monitoring and analysis of scientific and technical information in the field of physics and technology of memristive nanomaterials, fabrication of memristor-based electronic devices and neuromorphic systems (in particular, neurochips and neurointerfaces), the basic principles and approaches were suggested for the creation of the ANN prototype based on a matrix of 32 memristive elements, which is a two-layer perceptron, intended for solving non-linear classification tasks. The developed neural network uses as synaptic weights the technologically adapted and CMOS-compatible thin-film structures of the Au/Zr/oxide/TiN/Ti type and can be easily scaled down in the integrated version. Experimental study of electro-physical properties of memristive structures formed by using the methods of magnetron sputtering and computer simulation of microscopic physical-chemical processes of growth and rupture of conductive pathways (filaments) in the oxide material allows selecting optimal variants of structures of the Au/Zr/ZrO2(Y)/TiN/Ti and Au/Zr/SiO2/TiN/Ti types, the determination of dynamic ranges of resistance change and voltage values that correspond to an adequate (adaptive) change of resistance (conductivity). The scheme of synaptically coupled neuron-like generators was implemented for the first time by using the memristive devices, the characteristics of such a scheme were found. The resistance of memristive structure is a parameter characterizing the coupling strength. It is mandatory for the observation of different modes of synchronization that the resistance is changed gradually in a wide range of amplitudes of the master generator. The result is important to create the models of biosimilar neuromorphic systems imitating the training rules of neural networks of the brain, as well as to design and implement the next generation neurohybrid systems, where the memristive devices interact directly with the living neural network. The developed computer model of ANN is based on the proposed matrix of paired memristive elements (complementary pairs) in the ‘cross-point’ topology. Programming the weighting factor of the connection between electronic neurons is based on experimentally defined dependencies of the resistance of memristive elements on the amplitude of voltage pulse. The ANN performance was successfully demonstrated by the solution of classification tasks based on the type of input signal in the form of convex and concave functions in four samples. The weights obtained at this stage of training (resistance of memristor pairs) are regarded as the initial approximation for the second phase of training – direct hardware implementation of ANN model with the use of external control electronics. In order to create further an interface of ‘living neural network – memristor ANN’, the protocol of MEM cultivation of dissociated hippocampus cells was developed. The time of cultivation providing the most mature condition of synaptic contacts in the neural network of hippocampal cells, the cell density affecting the neural activity, the parameters of external stimulation of neural network required for the adequate response with plastic changes at the synaptic level were determined. The interface systems that are needed to process the bioelectric activity registered by MEM were selected and justified. Due to the limited size of the ANN prototype, it is proposed at this stage of the project to connect ANN inputs to selective electrodes, which most fully characterize the bioelectric activity of the living cultures of dissociated brain cells in response to a stimulus (pharmacological or electrical). To test the electrode selection approaches, the preliminary experiments were conducted on registration and stimulation of bioelectric activity of cell cultures, which allowed us to determine the criteria for finding statistically significant differences in the responses to the stimulation of different sites of neuronal culture. As a result of the project, an adaptive neurointerface of automatic registration and stimulation of living brain cells with a feedback should be demonstrated. This neurointerface can be used in the systems for monitoring of electrical activity of the central and peripheral nervous system, in the treatment of nervous system diseases or testing new pharmacological effects on the brain cells. It can be realized as a part of compact hybrid biochips and used in self-learning control systems of biocybernetic devices. In the longer perspective, in the case of construction of dense three-dimensional arrays of memristors, it will be possible to create electronic models of the human brain or its parts, what is important to restore the damaged areas of the central nervous system or to create biosimilar artificial intelligence. Thus, all the tasks planned in the first phase of the project are reached and create sufficient basis for the achievement of the main goal of the project. The comparison of the obtained results with the available scientific and technical information allows us to assess the level of research as relevant to the best world achievements in the field of creation of neuromorphic systems based on memristive nanomaterials. Information about the project on the Internet is available in the Research Gate profile of principal investigator: https://www.researchgate.net/profile/Alexey_Mikhaylov.

 

Publications

---

Annotation of the results obtained in 2017
During the reporting period, the cycle of research activities was continued and extended that provides the implementation of a neurointerface, which couples memristors or memristor-based artificial neural network (ANN) with the living neural network in a culture of dissociated hippocampal cells of the brain. The following activities were performed according to the plan: study of adaptive (synaptic) behavior of thin-film memristive structures and modification of their parameters by using the additional ion-beam processing; development and fabrication of ANN prototype based on array of memristors in a "cross-point" topology coupled with external (with respect to memristors) electronics; development of a circuit-level computer model of the cell, control methods involved in the ANN model, as well as its interaction with electronic interfaces; development of technologies for the transfer of registered signals of neuronal activity in the culture of dissociated hippocampal cells to the destination node of memristor-based ANN. Continuous monitoring and generalization of scientific and technical information on the project theme made it possible to reveal timely the important trends in the field, to elaborate tactical approaches to reach the specific project objectives and, what is most important, to outline strategy of development and application of the project results. It is concluded that the key to successful development and sustainability of the project is the balance in the combination of different approaches. The first (and the main) of these approaches is to demonstrate the potential of the "traditional" ANN in the form of a two-layer perceptron based on programmable memristive elements. The key advantages of the ANN being developed include, first of all, its multilayer structure, and hence the ability to solve nonlinear classification problems (based on the shape of input signal), which is very important when dealing with complex bioelectric activity, and secondly, the hardware implementation of all artificial network elements on one board, including the memristive synaptic chip, control electronics and neuron circuits. In the future, this arrangement will allow us to implement the adaptive neural interface "living neural network – memristor-based ANN" in the form of a compact autonomous device. The second approach that is pursued in parallel in the frame of the project is to find some alternative solutions for creating non-traditional neural network architectures where the stochastic nature and the "live" dynamics of memristive devices play a key role. These features of memristors make it possible to use them for direct processing and analysis of nerve cell activity, as well as for developing plausible physical models of spiking neural networks with self-organization of memristive connections between neurons. In the framework of the first approach, the design and technological solutions for the creation of an array of memristive elements in the "cross-point" topology was developed and implemented by the fabrication of chips mounted in a standard 64-pin package with sufficient yield of memristive devices (at least 32 per a chip) to create a weight matrix and their integration with the control electronics in the fabricated ANN prototype (microcontroller, RS-232 interface, circuits for the formation of programming voltage, power stabilizers), which provides the generation of programming pulses and implementation of the ANN learning algorithm. The technique was elaborated for electroforming and programming the specified value of resistance (conductivity) of a memristor by applying voltage pulses of certain amplitude. The technique provides a wide range of resistance (from less than 1 kΩ to values more than 1 MΩ) and low programming voltage (1-5 V). The investigated memristive devices are characterized by certain variation of resistance in different resistive states – either cycle-to-cycle or device-to-device – related to the stochastic nature of microscopic processes responsible for resistive switching. In order to improve the reproducibility of parameters of memristive devices for the use in ANN as programmable weight coefficients, additional processing of working oxide with low-energy Xe ions was included in the developed technological route. It is found that ion irradiation has a positive effect on both the growth process of filaments (conductive channels in oxide) and their original structure and geometry, which improves the reproducibility of electroforming and low-resistance state parameters as well as widens the resistance range by reducing the currents in high-resistance state. The process of switching between the resistive states continues to be stochastic in nature, what is important for the application of memristive devices in chaotic spiking neural networks. At this stage, the possibility of learning the memristor-based ANN prototype was tested on the individual memristive devices fabricated on the chip as well as on the developed circuit-level computer model of ANN cell taking into account the experimentally determined variability of resistive states of memristive elements. The developed control methods were tested using the ANN computer model to demonstrate the ability to solve the claimed non-linear classification tasks for an example of recognition of activity directions simulated in artificial neuronal culture. The obtained results made it possible to offer the technology for the delivery of neuronal activity signals recorded in the culture of dissociated hippocampal cells to the destination node – memristor-based ANN, which is based on a set of interrelated solutions on the fabrication of registration system, planting and cultivation of cell cultures, stimulation, registration and processing of neuronal activity, coding and delivery of signals to the ANN inputs. In the framework of the second approach, the experiments on the effect on a memristor of complex-waveform pulse sequences (including spike-like signals from the FitzHugh–Nagumo (FHN) neural generators) and the signals of neuronal activity directly recorded in the culture of brain cells on multi-electrode array) were taken to a new level. To register the real-time adaptive response of memristive devices to periodic signals (similar to potentiation and depression), the measurement system based on USB DAQ devices was designed and used. The obtained dependencies of resistance of memristive devices on the parameters of periodic signals were used to calibrate a mathematical model of synaptically coupled FHN neural generators for the first time developed in the frame of the project. This model is implemented as a single computer program that contains blocks simulating the dynamics of pre- and postsynaptic artificial FHN neurons and a block modeling processes in the memristive device. The model allows to predict the synchronization modes and to simulate more complex network architectures based on FHN generators and memristive devices. The experiments show that the resistance of memristive device is subject to both short-term (volatile), and long-term changes in response to signals of neuronal activity. This reaction depends on the initial state of memristor, parameters of spiking activity and allows offering new methods of real time processing of neuronal activity using direct coupling of memristors and living brain neurons. All the tasks planned in the second phase of the project are reached. The results obtained make an important contribution to the achievement of project goal and lay a groundwork for the transition to a qualitatively new level in the field of bio-similar memristive systems. The comparison of the obtained results with the available scientific and technical information allows us to assess the level of research as relevant to the best world achievements in the field of creation of neuromorphic systems based on memristive nanomaterials. As a result of the project, an adaptive neurointerface of automatic registration and stimulation of living brain cells with a feedback should be demonstrated. This neurointerface can be used in the systems for monitoring of electrical activity of the central and peripheral nervous system, in the treatment of nervous system diseases or testing new pharmacological effects on the brain cells. It can be realized as a part of compact hybrid biochips and used in self-learning control systems of biocybernetic devices. In the longer perspective, in the case of construction of dense three-dimensional arrays of memristors, it will be possible to create electronic models of the human brain or its parts, what is important to restore the damaged areas of the central nervous system or to create biosimilar artificial intelligence. Information about the project on the Internet is available in the Research Gate profile of principal investigator: https://www.researchgate.net/project/Biosimilar-memristive-systems-coupled-with-neuronal-biocultures.

 

Publications

---