ABSTRACT
A hybrid learning method of a software-based backpropagation learning and a hardware-based RWC learning is proposed for the development of circuit-based neural networks. The backpropagation is known as one of the most efficient learning algorithms. A weak point is that its hardware implementation is extremely difficult. The RWC algorithm, which is very easy to implement with respect to its hardware circuits, takes too many iterations for learning. The proposed learning algorithm is a hybrid one of these two. The main learning is performed with a software version of the BP algorithm, firstly, and then, learned weights are transplanted on a hardware version of a neural circuit. At the time of the weight transplantation, a significant amount of output error would occur due to the characteristic difference between the software and the hardware. In the proposed method, such error is reduced via a complementary learning of the RWC algorithm, which is implemented in a simple hardware. The usefulness of the proposed hybrid learning system is verified via simulations upon several classical learning problems.
ABSTRACT
Phase Change Memory (PCM) is an emerging technology exploiting the rapid and reversible phase transition of certain chalcogenides to realize nanoscale memory elements. PCM devices are being explored as non-volatile storage-class memory and as computing elements for in-memory and neuromorphic computing. It is well-known that PCM exhibits several characteristics of a memristive device. In this work, based on the essential physical attributes of PCM devices, we exploit the concept of Dynamic Route Map (DRM) to capture the complex physics underlying these devices to describe them as memristive devices defined by a state-dependent Ohm's law. The efficacy of the DRM has been proven by comparing numerical results with experimental data obtained on PCM devices.
ABSTRACT
Local activity is the capability of a system to amplify infinitesimal fluctuations in energy. Complex phenomena, including the generation of action potentials in neuronal axon membranes, may never emerge in an open system unless some of its constitutive elements operate in a locally active regime. As a result, the recent discovery of solid-state volatile memory devices, which, biased through appropriate DC sources, may enter a local activity domain, and, most importantly, the associated stable yet excitable sub-domain, referred to as edge of chaos, which is where the seed of complexity is actually planted, is of great appeal to the neuromorphic engineering community. This paper applies fundamentals from the theory of local activity to an accurate model of a niobium oxide volatile resistance switching memory to derive the conditions necessary to bias the device in the local activity regime. This allows to partition the entire design parameter space into three domains, where the threshold switch is locally passive (LP), locally active but unstable, and both locally active and stable, respectively. The final part of the article is devoted to point out the extent by which the response of the volatile memristor to quasi-static excitations may differ from its dynamics under DC stress. Reporting experimental measurements, which validate the theoretical predictions, this work clearly demonstrates how invaluable is non-linear system theory for the acquirement of a comprehensive picture of the dynamics of highly non-linear devices, which is an essential prerequisite for a conscious and systematic approach to the design of robust neuromorphic electronics. Given that, as recently proved, the potassium and sodium ion channels in biological axon membranes are locally active memristors, the physical realization of novel artificial neural networks, capable to reproduce the functionalities of the human brain more closely than state-of-the-art purely CMOS hardware architectures, should not leave aside the adoption of resistance switching memories, which, under the appropriate provision of energy, are capable to amplify the small signal, such as the niobium dioxide micro-scale device from NaMLab, chosen as object of theoretical and experimental study in this work.
ABSTRACT
Resistive Random Access Memories (RRAMs) are based on resistive switching (RS) operation and exhibit a set of technological features that make them ideal candidates for applications related to non-volatile memories, neuromorphic computing and hardware cryptography. For the full industrial development of these devices different simulation tools and compact models are needed in order to allow computer-aided design, both at the device and circuit levels. Most of the different RRAM models presented so far in the literature deal with temperature effects since the physical mechanisms behind RS are thermally activated; therefore, an exhaustive description of these effects is essential. As far as we know, no revision papers on thermal models have been published yet; and that is why we deal with this issue here. Using the heat equation as the starting point, we describe the details of its numerical solution for a conventional RRAM structure and, later on, present models of different complexity to integrate thermal effects in complete compact models that account for the kinetics of the chemical reactions behind resistive switching and the current calculation. In particular, we have accounted for different conductive filament geometries, operation regimes, filament lateral heat losses, the use of several temperatures to characterize each conductive filament, among other issues. A 3D numerical solution of the heat equation within a complete RRAM simulator was also taken into account. A general memristor model is also formulated accounting for temperature as one of the state variables to describe electron device operation. In addition, to widen the view from different perspectives, we deal with a thermal model contextualized within the quantum point contact formalism. In this manner, the temperature can be accounted for the description of quantum effects in the RRAM charge transport mechanisms. Finally, the thermometry of conducting filaments and the corresponding models considering different dielectric materials are tackled in depth.
ABSTRACT
Memristors represent the fourth electrical circuit element complementing resistors, capacitors and inductors. Hallmarks of memristive behavior include pinched and frequency-dependent I-V hysteresis loops and most importantly a functional dependence of the magnetic flux passing through an ideal memristor on its electrical charge. Microtubules (MTs), cylindrical protein polymers composed of tubulin dimers are key components of the cytoskeleton. They have been shown to increase solution's ionic conductance and re-orient in the presence of electric fields. It has been hypothesized that MTs also possess intrinsic capacitive and inductive properties, leading to transistor-like behavior. Here, we show a theoretical basis and experimental support for the assertion that MTs under specific circumstances behave consistently with the definition of a memristor. Their biophysical properties lead to pinched hysteretic current-voltage dependence as well a classic dependence of magnetic flux on electric charge. Based on the information about the structure of MTs we provide an estimate of their memristance. We discuss its significance for biology, especially neuroscience, and potential for nanotechnology applications.
Subject(s)
Electric Conductivity , Microtubules/metabolism , Biophysical Phenomena , Electric Impedance , Microtubules/chemistry , Nanotechnology , Neural Networks, Computer , Tubulin/chemistry , Tubulin/metabolismABSTRACT
This study firstly presents (i) a novel general cellular mapping scheme for two dimensional neuromorphic dynamical systems such as bio-inspired neuron models, and (ii) an efficient mixed analog-digital circuit, which can be conveniently implemented on a hybrid memristor-crossbar/CMOS platform, for hardware implementation of the scheme. This approach employs 4n memristors and no switch for implementing an n-cell system in comparison with 2n (2) memristors and 2n switches of a Cellular Memristive Dynamical System (CMDS). Moreover, this approach allows for dynamical variables with both analog and one-hot digital values opening a wide range of choices for interconnections and networking schemes. Dynamical response analyses show that this circuit exhibits various responses based on the underlying bifurcation scenarios which determine the main characteristics of the neuromorphic dynamical systems. Due to high programmability of the circuit, it can be applied to a variety of learning systems, real-time applications, and analytically indescribable dynamical systems. We simulate the FitzHugh-Nagumo (FHN), Adaptive Exponential (AdEx) integrate and fire, and Izhikevich neuron models on our platform, and investigate the dynamical behaviors of these circuits as case studies. Moreover, error analysis shows that our approach is suitably accurate. We also develop a simple hardware prototype for experimental demonstration of our approach.
ABSTRACT
Nanotechnology opens new ways to utilize recent discoveries in biological image processing by translating the underlying functional concepts into the design of CNN (cellular neural/nonlinear network)-based systems incorporating nanoelectronic devices. There is a natural intersection joining studies of retinal processing, spatio-temporal nonlinear dynamics embodied in CNN, and the possibility of miniaturizing the technology through nanotechnology. This intersection serves as the springboard for our multidisciplinary project. Biological feature and motion detectors map directly into the spatio-temporal dynamics of CNN for target recognition, image stabilization, and tracking. The neural interactions underlying color processing will drive the development of nanoscale multispectral sensor arrays for image fusion. Implementing such nanoscale sensors on a CNN platform will allow the implementation of device feedback control, a hallmark of biological sensory systems. These biologically inspired CNN subroutines are incorporated into the new world of analog-and-logic algorithms and software, containing also many other active-wave computing mechanisms, including nature-inspired (physics and chemistry) as well as PDE-based sophisticated spatio-temporal algorithms. Our goal is to design and develop several miniature prototype devices for target detection, navigation, tracking, and robotics. This paper presents an example illustrating the synergies emerging from the convergence of nanotechnology, biotechnology, and information and cognitive science.
Subject(s)
Artificial Intelligence , Biomimetics/instrumentation , Image Interpretation, Computer-Assisted/methods , Nanotechnology/instrumentation , Neural Networks, Computer , Retina/physiology , Transducers , Vision, Ocular/physiology , Animals , Biomimetics/methods , Equipment Design , Humans , Nanotechnology/methods , Pattern Recognition, Visual/physiology , Signal Processing, Computer-AssistedABSTRACT
Micro bimorph coils driven by a metalinsulator phase transition in VO2 function as powerful torsional muscles. Reversible torsional motion over one million cycles without degradation is demonstrated, with a superior rotational speed up to ca. 200,000 rpm, an amplitude of 500° per mm length, and a power density up to ca. 39 kW kg⻹.
ABSTRACT
We designed Adaptive Neuromorphic Architecture (ANA) that self-adjusts its inherent parameters (for instance, the resonant frequency) naturally following the stimuli frequency. Such an architecture is required for brain-like engineered systems because some parameters of the stimuli (for instance, the stimuli frequency) are not known in advance. Such adaptivity comes from a circuit element with memory, namely mem-inductor or mem-capacitor (memristor's sisters), which is history-dependent in its behavior. As a hardware model of biological systems, ANA can be used to adaptively reproduce the observed biological phenomena in amoebae.
Subject(s)
Models, Neurological , Neural Networks, Computer , Neurons/physiology , Neurosciences/instrumentation , Animals , Brain/cytology , HumansABSTRACT
Analog hardware architecture of a memristor bridge synapse-based multilayer neural network and its learning scheme is proposed. The use of memristor bridge synapse in the proposed architecture solves one of the major problems, regarding nonvolatile weight storage in analog neural network implementations. To compensate for the spatial nonuniformity and nonideal response of the memristor bridge synapse, a modified chip-in-the-loop learning scheme suitable for the proposed neural network architecture is also proposed. In the proposed method, the initial learning is conducted in software, and the behavior of the software-trained network is learned by the hardware network by learning each of the single-layered neurons of the network independently. The forward calculation of the single-layered neuron learning is implemented on circuit hardware, and followed by a weight updating phase assisted by a host computer. Unlike conventional chip-in-the-loop learning, the need for the readout of synaptic weights for calculating weight updates in each epoch is eliminated by virtue of the memristor bridge synapse and the proposed learning scheme. The hardware architecture along with the successful implementation of proposed learning on a three-bit parity network, and on a car detection network is also presented.