Hardware implementation of memristor-based SNN
Neural computation is assuming an increasing role for understanding the behavior of the brain and for emulating it in specific operations. The main goal in realizing analog VLSI systems able to mimic functionalities of biological neural networks is related to obtain a sufficient performance for the following features: 1) biological neuron input-output behavior 2) realistic and compact synapses.
About the first point, literature shows that reliable and precise latencies seem to have a basic role for the neural information representation. A very efficient neuron model exhibiting latency characteristics and a general good bio-plausibility was already presented by the authors in previous works. Now we propose a solution based on an analog circuit implementation. The resulting circuit shows a good matching with the ideal model and exhibits a relatively low complexity. The implementation uses a CMOS 90nm 6 Metal Copper low-K technology.
About the second point, we have to observe that in practical applications synapse/neuron ratio is very high, especially in the case of extremely dense networks (i.e., reproduction of a real scenario). For this reason synapses represent a considerable limitation in terms of waste of silicon area and power consumption. We propose a driving circuit model that does not require 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 (Spike-Timing-Dependent Plasticity) behavior allowing “controlled” changes for the synaptic weights.
For this research area, ELTLAB collaborates with DSPVLSI group of the same Department.
LIST OF PUBLICATIONS
CONFERENCE PAPERS
Resistive Switching Behaviour In ZnO-rGO composite thin film
Gian Carlo Cardarilli, Marco Re, Mario Salerno, Gianluca Susi, Alessandro Cristini, Simone Acciarito, Luca Di Nunzio, Rocco Fazzolari
2016 Technical Report ET2016. DOI:10.13140/RG.2.2.14030.87364
Hardware implementation of a spiking neuron model based on memristor
Simone Acciarito, Gian Carlo Cardarilli, Alessandro Cristini, Luca Di Nunzio, Rocco Fazzolari, Gaurav Mani Khanal, Marco Re, Gianluca Susi
2016 Technical Report GE2016.
Asynchronous Spiking Neural Networks
This research area concerns the development of an event-driven (or asynchronous) Spiking Neural Network (SNN) simulator, taking into account the most important biological neuron features, among which the spike latency (Fig. 1). Such neuronal feature represents a key factor for the information encoding. It has been conducted an accurate preliminary characterization of the spike latency by means of simulations carried out on the most accurate bio-realistic neuron model: the Hodgkin-Huxley model.
FIGURE 1. The red line indicates the spike latency as a function of the membrane potential Vm (red marked scale), or else of the current amplitude Iext, equivalently, obtained by means of simulations in NEURON environment. The dashed blue line indicates the the spike latency behavior, that we have approximated to a rectangular hyperbola. In addition, the normalized scale S (i.e., neuron state) is reported (blue marked scale). Note that, below the Sth (i.e., spiking threshold) value, no spike can be generated (fading blue area).
Further features have been implemented in the model, such as subthreshold decay, refractory period, inhibitory effect, synaptic plasticity, etc.
An ad-hoc event-driven method has been implemented in order to allow continuous-time simulations and preserve the information coding properties of the spike latency. This approach is characterized by high precision and high computational efficiency at the same time.
Currently, the simulator is implemented in Matlab, allowing asynchronous simulations of network composed by over 105 neurons. In the next future further improvements are planned: Java implementation, parallelization and hardware implementation.
It is also possible to apply some simple plasticity algorithms in order to emulate interesting global phenomena, such as the "Neuronal Group Selection" (Fig. 2) or the "jitter-reduction".
Figure 2. Formation of neuronal groups after a stimulation consisting of pseudo-random spike sequences.
Such a parallel systems can be used to handle the classic issues deal with artificial neural networks (e.g., pattern recognition), but also to model the biological neural circuitry.
AWARDS
Best Paper Category, received at the International Conference "ACEEE, ACE 2012" (Amsterdam, NL) for the presented paper "Spiking neural networks as analog dynamical systems: basic paradigm and simple applications" (M. Salerno, G. Susi, A. D’Annessa, A. Cristini, Y. Sanfelice - June 2012).
Best Poster Award, received at the 29th Annual Meeting of Circuit Theory Researchers (Padova University, IT), for the presented poster "Continuous-time Spiking Neural Networks: general paradigm and event-driven simulation". (M. Salerno, G. Susi, A. Cristini, M. Re, G.C. Cardarilli - June 2013).
LIST OF PUBLICATIONS
JOURNALS
Path multimodality in a Feedforward SNN module, using LIF with Latency model
Spiking Neural Networks As Continuous-Time Dynamical Systems: Fundamentals, Elementary Structures And Simple Applications
Mario Salerno, Gianluca Susi, Alessandro Cristini, Yari Sanfelice, Andrea D’Annessa
2013 ACEEE Int. J. on Information Technology, Vol. 3, No. 1, March 2013, pp. 80-89. ISSN:2158-0138, DOI:01.IJIT.3.1.1129
CONFERENCE PAPERS
Spiking Neural Networks based on LIF with Latency: Simulation and Synchronization Effects
Gian Carlo Cardarilli, Alessandro Cristini, Luca Di Nunzio, Marco Re, Mario Salerno, Gianluca Susi
2013 IEEE Asilomar Conference on Signals, Systems, and Computers, 3-6 Nov. 2013, Pacific Grove, CA, USA, pp. 1838-1842. ISBN:978-1-4799-2388-5, DOI:10.1109/ACSSC.2013.6810620. IEEE CONFERENCE PUBLICATIONS.
Spiking neural networks as analog dynamical systems: basic paradigm and simple applications
Mario Salerno, Gianluca Susi, Andrea D’Annessa, Alessandro Cristini, Yari Sanfelice
2012 Proc. of the 3rd Int. Conf. on advances in Computer Engineering, 2012, pp. 17-23. Publisher: ACEEE. ISBN:978-90-819067-8-4, DOI:02.ACE.2012.03.10
Accurate latency characterization for very large asynchronous spiking neural networks
Mario Salerno, Gianluca Susi, Alessandro Cristini
2011 Proc. of the 4th Int. Conf. on Bioinformatics Models, Methods and Algorithms, 2011, pp. 116-124. Publisher: SciTePress. ISBN:978-989-8425-36-2. DOI:10.5220/0003134601160124
BOOK CHAPTERS
A Continuous-Time Spiking Neural Network Paradigm
Alessandro Cristini, Mario Salerno, Gianluca Susi
2015 Advances in Neural Networks: Computational and Theoretical Issues, Smart Innovation, Systems and Technologies, Vol. 37, pp. 49-60, 2015. ISBN:978-3-319-18163-9, ISSN:2190-3018, DOI:10.1007/978-3-319-18164-6_6
POSTERS
Spiking neural network-based character recognition exploiting different coding schemes
Event-driven simulation of continuous-time neural networks
Mario Salerno, Gianluca Susi, Alessandro Cristini, Marco Re, Gian Carlo Cardarilli
2014 30th Riunione Annuale dei Ricercatori di Elettrotecnica (ET2014), Sorrento, IT. DOI:10.13140/2.1.4182.3045
Continuous-time Spiking Neural Networks: general paradigm and event-driven simulation
Mario Salerno, Gianluca Susi, Alessandro Cristini, Marco Re, Gian Carlo Cardarilli
2013 29th Riunione Annuale dei Ricercatori di Elettrotecnica (ET2013), Padova, IT. DOI:10.13140/2.1.2144.5443
TECHNICAL REPORTS
A simple approach for different-scale CTNN simulations
Mario Salerno, Gianluca Susi, Alessandro Cristini, Gian Carlo Cardarilli, Marco Re
Continuous-Time Neural Networks: paradigm and applications
Navigation aid for the blind
AudiNect Project: this project is focused on the implementation of a novel tool for the autonomous navigation of blind people, in order to overcome some criticalities of the classic ones, based on sonar, infrared, GPS or RFID technique. The tool proposed here exploits the sense of hearing. Note that, this sense is very developed in subjects with visual impairment. AudiNect can be considered as a "portable sonar", signaling by means of acoustic feedback the presence of obstacles along a particular path that an impaired person needs to avoid. In particular, the depth data are provided as a Depth Map using the Microsoft Kinect. Then, the data is processed in order to encode a mapping of the data into properly acoustic feedback (Fig. 1), easily understandable by the user, allowing to search for free routes.
The software is implemented using Processing, whereas the acoustic feedback is generated using PureData. Processing is able to interact with Kinect through Simple OpenNI library, and communicates by means of OSCp5 library.
FIGURE 1. Example of acoustic signals used for training the subject in order to test AudiNect system.
The system was tested on blindfolded sighted individuals. The related tests have produced satisfactory results on the basis of learning curves, showing a rapid adaptation of the individuals to the proposed method. This suggest that the technology is well integrated with the human cognitive processes. Indeed, AudiNect lets the user to easily identify the best free path.
BrailNect Project: this project represents a support for the above mentioned AudiNect (Fig. 2). The system takes information from the surrounding environment by means of a Kinect, and provides feedback using a dedicated haptic display (an electromechanical device).
FIGURE 2. Complete scheme of the whole system AudiNect - BrailNect.
LIST OF PUBLICATIONS
JOURNALS
AudiNect: An Aid for the Autonomous Navigation of Visually Impaired People, Based On Virtual Interface 
Link
Mario Salerno, Marco Re, Alessandro Cristini, Gianluca Susi, Marco Bertola, Emiliano Daddario, Francesca Capobianco
2013 Int. J. of Human Computer Interaction (IJHCI), Vol. 4, Issue 1, Feb. 2013, pp. 25-33. ISSN:2180-1347.
CONFERENCE PAPERS
A low-cost Indoor and Outdoor terrestrial autonomous navigation model
Gianluca Susi, Alessandro Cristini, Mario Salerno, Emiliano Daddario
2014 IEEE 22nd Telecommunications Forum TELFOR, 25-27 Nov 2014, Belgrade, Serbia, pp. 675-678. ISBN:978-1-4799-6190-0. DOI:10.1109/TELFOR.2014.7034499. IEEE CONFERENCE PUBLICATIONS.
Audio Engineering
ELTLAB and DSPVLSI group coordinate the Master in Audio Engineering of the University of Rome "Tor Vergata".
Research
- Asynchronous spiking neural networks
- Memristor-based neuromorphic hardware
- Navigation aid for the blind
- Audio Engineering
Research achievements:
- Best Paper Award @ 3rd International Conference on Advances in Computer Engineering (ACEEE 2012) - IDES;
- Best Poster award @ 29° Meeting of Electrical Engineering researchers (ET2013).