Multicasting mesh AER: a scalable assembly approach for reconfigurable neuromorphic structured AER systems. Application to ConvNets.

This paper presents a modular, scalable approach to assembling hierarchically structured neuromorphic Address Event Representation (AER) systems. The method consists of arranging modules in a 2D mesh, each communicating bidirectionally with all four neighbors. Address events include a module label. Each module includes an AER router which decides how to route address events. Two routing approaches have been proposed, analyzed and tested, using either destination or source module labels. Our analyses reveal that depending on traffic conditions and network topologies either one or the other approach may result in better performance. Experimental results are given after testing the approach using high-end Virtex-6 FPGAs. The approach is proposed for both single and multiple FPGAs, in which case a special bidirectional parallel-serial AER link with flow control is exploited, using the FPGA Rocket-I/O interfaces. Extensive test results are provided exploiting convolution modules of 64 × 64 pixels with kernels with sizes up to 11 × 11, which process real sensory data from a Dynamic Vision Sensor (DVS) retina. One single Virtex-6 FPGA can hold up to 64 of these convolution modules, which is equivalent to a neural network with 262 × 10(3) neurons and almost 32 million synapses.

STDP and STDP variations with memristors for spiking neuromorphic learning systems.

In this paper we review several ways of realizing asynchronous Spike-Timing-Dependent-Plasticity (STDP) using memristors as synapses. Our focus is on how to use individual memristors to implement synaptic weight multiplications, in a way such that it is not necessary to (a) introduce global synchronization and (b) to separate memristor learning phases from memristor performing phases. In the approaches described, neurons fire spikes asynchronously when they wish and memristive synapses perform computation and learn at their own pace, as it happens in biological neural systems. We distinguish between two different memristor physics, depending on whether they respond to the original "moving wall" or to the "filament creation and annihilation" models. Independent of the memristor physics, we discuss two different types of STDP rules that can be implemented with memristors: either the pure timing-based rule that takes into account the arrival time of the spikes from the pre- and the post-synaptic neurons, or a hybrid rule that takes into account only the timing of pre-synaptic spikes and the membrane potential and other state variables of the post-synaptic neuron. We show how to implement these rules in cross-bar architectures that comprise massive arrays of memristors, and we discuss applications for artificial vision.

Compact low-power calibration mini-DACs for neural arrays with programmable weights.

This paper considers the viability of compact low-resolution low-power mini digital-to-analog converters (mini-DACs) for use in large arrays of neural type cells, where programmable weights are required. Transistors are biased in weak inversion in order to yield small currents and low power consumptions, a necessity when building large size arrays. One important drawback of weak inversion operation is poor matching between transistors. The resulting effective precision of a fabricated array of 50 DACs turned out to be 47% (1.1 bits), due to transistor mismatch. However, it is possible to combine them two by two in order to build calibrated DACs, thus compensating for inter-DAC mismatch. It is shown experimentally that the precision can be improved easily by a factor of 10 (4.8% or 4.4 bits), which makes these DACs viable for low-resolution applications such as massive arrays of neural processing circuits. A design methodology is provided, and illustrated through examples, to obtain calibrated mini-DACs of a given target precision. As an example application, we show simulation results of using this technique to calibrate an array of digitally controlled integrate-and-fire neurons.

Log-domain implementation of complex dynamics reaction-diffusion neural networks.

We have identified a second-order reaction-diffusion differential equation able to reproduce through parameter setting different complex spatio-temporal behaviors. We have designed a log-domain hardware that implements the spatially discretized version of the selected reaction-diffusion equation. The logarithmic compression of the state variables allows several decades of variation of these state variables within subthreshold operation of the MOS transistors. Furthermore, as all the equation parameters are implemented as currents, they can be adjusted several decades. As a demonstrator, we have designed a chip containing a linear array of ten second-order dynamics coupled cells. Using this hardware, we have experimentally reproduced two complex spatio-temporal phenomena: the propagation of travelling waves and of trigger waves, as well as isolated oscillatory cells.

A programmable VLSI filter architecture for application in real-time vision processing systems.

An architecture is proposed for the realization of real-time edge-extraction filtering operation in an Address-Event-Representation (AER) vision system. Furthermore, the approach is valid for any 2D filtering operation as long as the convolutional kernel F(p,q) is decomposable into an x-axis and a y-axis component, i.e. F(p,q)=H(p)V(q), for some rotated coordinate system [p,q]. If it is possible to find a coordinate system [p,q], rotated with respect to the absolute coordinate system a certain angle, for which the above decomposition is possible, then the proposed architecture is able to perform the filtering operation for any angle we would like the kernel to be rotated. This is achieved by taking advantage of the AER and manipulating the addresses in real time. The proposed architecture, however, requires one approximation: the product operation between the horizontal component H(p) and vertical component V(q) should be able to be approximated by a signed minimum operation without significant performance degradation. It is shown that for edge-extraction applications this filter does not produce performance degradation. The proposed architecture is intended to be used in a complete vision system known as the Boundary-Contour-System and Feature-Contour-System Vision Model, proposed by Grossberg and collaborators. The present paper proposes the architecture, provides a circuit implementation using MOS transistors operated in weak inversion, and shows behavioral simulation results at the system level operation and electrical simulation and experimental results at the circuit level operation of some critical subcircuits.

An ART1 microchip and its use in multi-ART1 systems.

Recently, a real-time clustering microchip neural engine based on the ART1 architecture has been reported. However, that chip rendered an extremely high silicon area consumption of 1 cm(2), and consequently an extremely low yield of 6%. Redundant circuit techniques can be introduced to improve yield performance at the cost of further increasing chip size. In this paper we present an improved ART1 chip prototype based on a different approach to implement the most area consuming circuit elements of the first prototype: an array of several thousand current sources which have to match within a precision of around 1%. Such achievement was possible after a careful transistor mismatch characterization of the fabrication process (ES2-1.0 mum CMOS). A new prototype chip has been fabricated which can cluster 50-b input patterns into up to ten categories. The chip has 15 times less area, shows a yield performance of 98%, and presents the same precision and speed than the previous prototype. Due to its higher robustness multichip systems are easily assembled. As a demonstration we show results of a two-chip ART1 system, and of an ARTMAP system made of two ART1 chips and an extra interfacing chip.

A CMOS analog adaptive BAM with on-chip learning and weight refreshing.

The transconductance-mode (T-mode) approach is extended to implement analog continuous-time neural network hardware systems to include on-chip Hebbian learning and on-chip analog weight storage capability. The demonstration vehicle used is a 5+5-neuron bidirectional associative memory (BAM) prototype fabricated in a standard 2-mum double-metal double-polysilicon CMOS process. Mismatches and nonidealities in learning neural hardware are not supposed to be critical if on-chip learning is available, because they will be implicitly compensated. However, mismatches in the learning circuits themselves cannot always be compensated. This mismatch is specially important if the learning circuits use transistors operating in weak inversion. The authors estimate the expected mismatch between learning circuits in the BAM network prototype and evaluate its effect on the learning performance, using theoretical computations and Monte Carlo HSPICE simulations. These theoretical predictions are verified using experimentally measured results on the test vehicle prototype.