This paper introduces a novel methodology for training an event-driven classifier within a Spiking Neural Network (SNN) System capable of yielding good classification results when using both synthetic input data and real data captured from Dynamic Vision Sensor (DVS) chips. The proposed supervised method uses the spiking activity provided by an arbitrary topology of prior SNN layers to build histograms and train the classifier in the frame domain using the stochastic gradient descent algorithm. In addition, this approach can cope with leaky integrate-and-fire neuron models within the SNN, a desirable feature for real-world SNN applications, where neural activation must fade away after some time in the absence of inputs. Consequently, this way of building histograms captures the dynamics of spikes immediately before the classifier. We tested our method on the MNIST data set using different synthetic encodings and real DVS sensory data sets such as N-MNIST, MNIST-DVS, and Poker-DVS using the same network topology and feature maps. We demonstrate the effectiveness of our approach by achieving the highest classification accuracy reported on the N-MNIST (97.77%) and Poker-DVS (100%) real DVS data sets to date with a spiking convolutional network. Moreover, by using the proposed method we were able to retrain the output layer of a previously reported spiking neural network and increase its performance by 2%, suggesting that the proposed classifier can be used as the output layer in works where features are extracted using unsupervised spike-based learning methods. In addition, we also analyze SNN performance figures such as total event activity and network latencies, which are relevant for eventual hardware implementations. In summary, the paper aggregates unsupervised-trained SNNs with a supervised-trained SNN classifier, combining and applying them to heterogeneous sets of benchmarks, both synthetic and from real DVS chips.

Journal Paper - Frontiers in Neuroscience, vol. 11, article 350, 2017 FRONTIERS RESEARCH FOUNDATION
DOI: 10.3389/fnins.2017.00350    ISSN: 1662-4548    » doi

In recent years, the application of evolutionary computation techniques to electronic circuit design problems, ranging from digital to analog and radiofrequency circuits, has received increasing attention. The level of maturity runs inversely to the complexity of the design task, less complex in digital circuits, higher in analog ones and still higher in radiofrequency circuits. Radiofrequency inductors are key culprits of such complexity. Their key performance parameters are inductance and quality factors, both a function of the frequency. The inductor optimization requires knowledge of such parameters at a few representative frequencies. Most common approaches for optimization-based radiofrequency circuit design use analytical models for the inductors. Although a lot of effort has been devoted to improve the accuracy of such analytical models, errors in inductance and quality factor in the range of 5%-25% are usual and it may go as high as 200% for some device sizes. When the analytical models are used in optimization-based circuit design approaches, these errors lead to suboptimal results, or, worse, to a disastrous non-fulfilment of specifications. Expert inductor designers rely on iterative evaluations with electromagnetic simulators, which, properly configured, are able to yield a highly accurate performance evaluation. Unfortunately, electromagnetic simulations typically take from some tens of seconds to a few hours, hampering their coupling to evolutionary computation algorithms. Therefore, analytical models and electromagnetic simulation represent extreme cases of the accuracy-efficiency trade-off in performance evaluation of radiofrequency inductors. Surrogate modeling strategies arise as promising candidates to improve such trade-off. However, obtaining the necessary accuracy is not that easy as inductance and quality factor at some representative frequencies must be obtained and both performances change abruptly around the self-resonance frequency, which is particular to each device and may be located above or below the frequencies of interest. Both, offline and online training methods will be considered in this work and a new two-step strategy for inductor modeling is proposed that significantly improves the accuracy of offline methods The new strategy is demonstrated and compared for both, single-objective and multi-objective optimization scenarios. Numerous experimental results show that the proposed two-step approach outperforms simpler application strategies of surrogate modelling techniques, getting comparable performances to approaches based on electromagnetic simulation but with orders of magnitude less computational effort.

Journal Paper - Applied Soft Computing, vol. 60, pp 495-507, 2017 ELSEVIER
DOI: 10.1016/j.asoc.2017.07.036    ISSN: 1568-4946    » doi

Networks with a modular structure are expected to have a lower risk of global failure. However, this theoretical result has remained untested until now. We used an experimental microarthropod metapopulation to test the effect of modularity on the response to perturbation. We perturbed one local population and measured the spread of the impact of this perturbation, both within and between modules. Our results show the buffering capacity of modular networks. To assess the generality of our findings, we then analyzed a dynamical model of our system. We show that in the absence of perturbations, modularity is negatively correlated with metapopulation size. However, even when a small local perturbation occurs, this negative effect is offset by a buffering effect that protects the majority of the nodes from the perturbation.

Journal Paper - Science, vol. 357, no. 6347, pp 199-201, 2017 AAAS
DOI: 10.1126/science.aal4122    ISSN: 0036-8075    » doi

The design of a direct time-of-flight complementary metal-oxide-semiconductor (CMOS) image sensor (dToF-CIS) based on a single-photon avalanche-diode (SPAD) array with an in-pixel time-to-digital converter (TDC) must contemplate system-level aspects that affect its overall performance. This paper provides a detailed analysis of the impact of process parameters, voltage supply, and temperature (PVT) variations on the time bin of the TDC array. Moreover, the design and characterization of a global compensation loop is presented. It is based on a phase locked loop (PLL) that is integrated on-chip. The main building block of the PLL is a voltage-controlled ring-oscillator (VCRO) that is identical to the ones employed for the in-pixel TDCs. The reference voltage that drives the master VCRO is distributed to the voltage control inputs of the slave VCROs such that their multiphase outputs become invariant to PVT changes. These outputs act as time interpolators for the TDCs. Therefore the compensation scheme prevents the time bin of the TDCs from drifting over time due to the aforementioned factors. Moreover, the same scheme is used to program different time resolutions of the direct time-of-flight (ToF) imager aimed at 3D ranging or depth map imaging. Experimental results that validate the analysis are provided as well. The compensation loop proves to be remarkably effective. The spreading of the TDCs time bin is lowered from: (i) 20% down to 2.4% while the temperature ranges from 0 °C to 100 °C; (ii) 27% down to 0.27%, when the voltage supply changes within ±10% of the nominal value; (iii) 5.2 ps to 2 ps standard deviation over 30 sample chips, due to process parameters´ variation.

Journal Paper - Sensors, vol. 17, no. 5, art. 1072, 2017 MDPI
DOI: 10.3390/s17051072    ISSN: 1424-8220    » doi

We report a high dynamic range (HDR) image sensor with a linear response that overcomes some of the limitations of sensors with pixels with self-reset operation. It operates similar to an active pixel sensor, but its pixels have a novel asynchronous event-based overflow detection mechanism. Whenever the pixel voltages at the integration capacitance reach a programmable threshold, the pixels self-reset and send out asynchronously an event indicating this. At the end of the integration period, the voltage at the integration capacitance is digitized and readout. Combining this information with the number of events fired by each pixel, it is possible to render linear HDR images. Event operation is transparent to the final user. There is no limitation for the number of self-resets of each pixel. The output data format is compatible with frame-based devices. The sensor was fabricated in the AMS 0.18-μm HV technology. A detailed system description and experimental results are provided in this paper. The sensor can render images with an intra-scene dynamic range of up to 130 dB with linear outputs. The pixels' pitch is 25 μm and the sensor power consumption is 58.6 mW.

Journal Paper - IEEE Journal of Solid-State Circuits, vol. 52, no. 6, pp 1605-1617, 2017 IEEE
DOI: 10.1109/JSSC.2017.2679058    ISSN: 0018-9200    » doi