Gianluca Susi

ACCURATE LATENCY CHARACTERIZATION FOR VERY LARGE ASYNCHRONOUS SPIKING NEURAL NETWORKS

The simulation problem of very large fully asynchronous Spiking Neural Networks is considered in this paper. To this purpose, a preliminary accurate analysis of the latency time is made, applying classical modelling methods to single neurons. The latency characterization is then used to propose a simplified model, able to simulate large neural networks. On this basis, networks, with up to 100,000 neurons for more than 100,000 spikes, can be simulated in a quite short time with a simple MATLAB program. Plasticity algorithms are also applied to emulate interesting global effects as the Neuronal Group Selection.

Spiking neural networks based on LIF with latency: Simulation and synchronization effects

In this paper, a work on spiking neural networks based on a model of a kind of Leaky Integrate-and-Fire (LIF) neuron with latency is presented. Efficient simulations are carried out through an ad hoc event-driven approach, highlighting some particular effects of synchrony in a simple feedforward network topology. These results are consistent with literature results and, thanks to the implementation of the biologically plausible latency effect in the model, new results have emerged from the simulations. The authors plan to apply these results in the near future to applications in which this kind of neural networks and Digital Signal Processing (DSP) applications can be merged to obtain powerful nonlinear DSP techniques. In the plan of the authors is also the definition of a hardware prototype of the network based on analog/digital techniques.

Hardware design of LIF with Latency neuron model with memristive STDP synapses

In this paper, the hardware implementation of a neuromorphic system is presented. This system is composed of a Leaky Integrate-and-Fire with Latency (LIFL) neuron and a Spike-Timing Dependent Plasticity (STDP) synapse. LIFL neuron model allows to encode more information than the common Integrate-and-Fire models, typically considered for neuromorphic implementations. In our system LIFL neuron is implemented using CMOS circuits while memristor is used for the implementation of the STDP synapse. A description of the entire circuit is provided. Finally, the capabilities of the proposed architecture have been evaluated by simulating a motif composed of three neurons and two synapses. The simulation results confirm the validity of the proposed system and its suitability for the design of more complex spiking neural networks

A spiking neural network-based model for anaerobic digestion process

There are many conversion technologies for the transformation of biomass into usable energy forms. Among these technologies, anaerobic digestion is one of the most attractive. In many papers appeared in the literature it has been demonstrated that the application of efficient mathematical models is an essential requirement to improve digesters performance. In this paper a spiking neural network-based model for anaerobic digestion process is proposed. This model performs a long-term prediction of the concentration of the biogas (CH4 and CO2) at the 100th day of the process, by analysing the concentration evolution of 6 measurable marker-molecules (MMM) namely CH4, CH4S, CO2, H2, H2S and NH3 during the first 10 days of the process. For the validation of the model, a small domestic digester was realized. The tests carried out show an excellent agreement between the predicted values and those obtained with the digester.

An a VLSI driving circuit for memristor-based STDP

The main goal in realizing a VLSI (analog VLSI) systems able to mimic functionalities of biological neural networks is pointed to the reproduction of realistic synapses. Indeed, because of the relative high synapse/neuron ratio, especially in the case of extremely dense networks (i.e., reproduction of a real scenario), synapses represent a considerable limitation in terms of waste of silicon area and power consumption as well. Thanks to advancement made in the implementation of memristor, the interest in bio-inspired neural network design has been renewed. Memristors have tunable resistance which depends on its past state; this is analogous to the operating mode of biological synapses. In this paper, we present the circuit implementation of a simple memristor-based neural network. Here, we propose a driving circuit model that not requires specific shape input pulses to change the memristor conductance (i.e., synaptic strength), but it can be driven by arbitrary shaped input pulses. Moreover, this prototype circuit offers the chance of emulating the standard STDP behavior allowing “controlled” changes for the synaptic weights. Some preliminary experimental results are reported to validate the proposed driving circuit.

A Continuous-Time Spiking Neural Network Paradigm

In this work, a novel continuous-time spiking neural network paradigm is presented. Indeed, because of a neuron can fire at any given time, this kind of approach is necessary. For the purpose of developing a simulation tool having such a property, an ad-hoc event-driven method is implemented. A simplified neuron model is introduced with characteristics similar to the classic Leaky Integrate-and-Fire model, but including the spike latency effect. The latency takes into account that the firing of a given neuron is not instantaneous, but occurs after a continuous-time delay. Both excitatory and inhibitory neurons are considered, and simple synaptic plasticity rules are modeled. Nevetheless the chance to customize the network topology, an example with Cellular Neural Network (CNN)-like connections is presented, and some interesting global effects emerging from the simulations are reported.

A low-cost indoor and outdoor terrestrial autonomous navigation model

In this paper, a method for low-cost system design oriented to indoor and outdoor autonomous navigation is illustrated. In order to provide a motivation for the solution here presented, a brief discussion of the typical drawbacks of state-of-the-art technologies is reported. Finally, an application of such a method for the design of a navigation system for blindfolded people is shown.

Path multimodality in a feedforward SNN module, using LIF with latency model

In this paper, the network transmission properties of a feedforward Spiking Neural Network (SNN) affected by synchronous stimuli are investigated with respect to the connection probability and the synaptic strengths. By means of an event-driven method, all simulations are conducted using the Leaky Integrateand-Fire with Latency (LIFL) model. Typical cases are taken into consideration, in which a network section (module) is able to process the input information, introducing a particular behavior, that we have called path multimodality. Simulation results are discussed. Through this phenomenon, the output layer of the network can generate a number of temporally spaced groups of synchronous spikes. The multimodality effect could be applied for various purposes, for instance in coding or else transmission issues

ZnO-rGO composite thin film resistive switching device: emulating biological synapse behavior

We have fabricated Sol-Gel synthesised Zinc Oxide (ZnO)-Reduced Graphene Oxide (rGO) on Fluorine-doped tin oxide (FTO) glass electrodes using a Dip Coating process. The Ag/ZnO-rGO/FTO sandwich structure showed bipolar resistive switching behavior. The resistive switching behavior can be attributed to the oxygen vacancies in the ZnO-rGO composite thin film giving rise to the formation and annihilation of conducting filament along the thin film. Good resistive switching (RS) characteristics with good On-OFF was also observed with good stability. The fabricated device has characteristics similar to that of biological synaptic plasticity and can be used for making electronics dynamical synapse.

Event-driven simulation of continuous-time neural networks

The transforming of incoming signals into action potentials by neurons is believed to be the basis for information processing in nervous systems. In many cases, the accurate representation of involved timings variability is necessary for a correct computation in neural network simulations. A lot of nervous system simulations reported in scientific literature are computed with time-step based methods. This technique is valid to describe many aspects of biological circuits, but some computational aspects (inefficiency, unreliability, etc.) have been highlighted when used in certain scenarios, especially on very large nets. In this work, a very simple and effective analog spiking neural network simulator, based on LIF (Leaky Integrate and Fire) with latency neurons, is presented. It is simulated with an event-driven method, necessary to guarantee the preservation of the original process behavior. In this way, the simulation proceeds without any forcing in order to obtain a compromise between high precision and computational cost. Networks with up to 105 neurons for more than 105 spikes, can be simulated in a few minutes (using a standard PC) with a simple MATLAB tool. Plasticity algorithms are also applied to develop bio-inspired applications and emulate interesting global effects as the Neuronal Group Selection.