R. Gadea

Artificial neural network implementation on a single FPGA of a pipelined on-line backpropagation

The paper describes the implementation of a systolic array for a multilayer perceptron on a Virtex XCV400 FPGA with a hardware-friendly learning algorithm. A pipelined adaptation of the on-line backpropagation algorithm is shown. Parallelism is better exploited because both forward and backward phases can be performed simultaneously. We can implement very large interconnection layers by using large Xilinx devices with embedded memories alongside the projection used in the systolic architecture. These physical and architectural features --- together with the combination of FPGA reconfiguration properties with a design flow based on generic VHDL --- create an easy, flexible, and fast method of designing a complete ANN on a single FPGA. The result offers a high degree of parallelism and fast performance.

Octree-based, GPU implementation of a continuous cellular automaton for the simulation of complex, evolving surfaces

Abstract Presently, dynamic surface-based models are required to contain increasingly larger numbers of points and to propagate them over longer time periods. For large numbers of surface points, the octree data structure can be used as a balance between low memory occupation and relatively rapid access to the stored data. For evolution rules that depend on neighborhood states, extended simulation periods can be obtained by using simplified atomistic propagation models, such as the Cellular Automata (CA). This method, however, has an intrinsic parallel updating nature and the corresponding simulations are highly inefficient when performed on classical Central Processing Units (CPUs), which are designed for the sequential execution of tasks. In this paper, a series of guidelines is presented for the efficient adaptation of octree-based, CA simulations of complex, evolving surfaces into massively parallel computing hardware. A Graphics Processing Unit (GPU) is used as a cost-efficient example of the parallel architectures. For the actual simulations, we consider the surface propagation during anisotropic wet chemical etching of silicon as a computationally challenging process with a wide-spread use in microengineering applications. A continuous CA model that is intrinsically parallel in nature is used for the time evolution. Our study strongly indicates that parallel computations of dynamically evolving surfaces simulated using CA methods are significantly benefited by the incorporation of octrees as support data structures, substantially decreasing the overall computational time and memory usage.

PESIC: An Integrated Front-End for PET Applications

An ASIC front-end has been developed for multi-anode photomultiplier based nuclear imaging devices. Its architecture has been designed to improve resolution and decrease pile-up probability in Positron Emission Tomography systems which employ continuous scintillator crystals. Analog computation elements are isolated from the photomultiplier by means of a current sensitive preamplifier stage. This allows digitally programmable adjustment of every anode gain, also providing better resolution in gamma event position calculation and a shorter front-end deadtime. The preamplifier stage also offers the possibility of using other types of photomultiplier devices such as SiPM. The ASIC architecture includes measurement of the depth of interaction of the gamma event based on the width of the light distribution in order to reduce parallax error and increase spatial resolution during image reconstruction stage. An output stage of transresistance amplifiers offer voltage output signals which may be introduced in the A/D conversion stage with no further processing.

Corrected position estimation in PET detector modules with multi-anode PMTs using neural networks

This paper studies the use of Neural Networks (NNs) for estimating the position of impinging photons in gamma ray detector modules for PET cameras based on continuous scintillators and Multi-Anode Photomultiplier Tubes (MA-PMTs). The detector under study is composed of a 49/spl times/49/spl times/10 mm/sup 3/ continuous slab of LSO coupled to a flat panel H8500 MA-PMT. Four digitized signals from a charge division circuit, which collects currents from the 8/spl times/8 anode matrix of the photomultiplier, are used as inputs to the NN, thus reducing drastically the number of electronic channels required. We have simulated the computation of the position for 511 keV gamma photons impacting perpendicularly to the detector surface. Thus, we have performed a thorough analysis of the NN architecture and training procedures in order to achieve the best results in terms of spatial resolution and bias correction. Results obtained using GEANT4 simulation toolkit show a resolution of 1.3 mm/1.9 mm FWHM at the center/edge of the detector and less than 1 mm of systematic error in the position near the edges of the scintillator. The results confirm that NNs can partially model and correct the non-uniform detector response using only the position-weighted signals from a simple 2D DPC circuit. Linearity degradation for oblique incidence is also investigated. Finally, the NN can be implemented in hardware for parallel real time corrected Line-of-Response (LOR) estimation. Results on resources occupancy and throughput in FPGA are presented.

Predictive assessment of ochratoxin A accumulation in grape juice based-medium by Aspergillus carbonarius using neural networks.

Aims:  To study the ability of multi‐layer perceptron artificial neural networks (MLP‐ANN) and radial‐basis function networks (RBFNs) to predict ochratoxin A (OTA) concentration over time in grape‐based cultures of Aspergillus carbonarius under different conditions of temperature, water activity (aw) and sub‐inhibitory doses of the fungicide carbendazim.

DOI measurement with monolithic scintillation crystals: A primary performance evaluation

We report a first assessment of image quality enhancement achieved by the implementation of depth of interaction detection with monolithic crystals. The method of interaction depth measurement is based on analogue computation of the standard deviation with an enhanced charge divider readout. This technique of depth of interaction detection was developed in order to provide fast and determination of this parameter at a reasonable increase of detector cost. The detector consists of an large-sized monolithic scintillator coupled to a position sensitive photomultiplier tube. A special design feature is the flat-topped pyramidal shape of the crystal. This reduces image compression near the edges of the scintillator. We studied the image enhancement qualitatively with a FDG filled hot spot phantom and quantitatively by displacing a single point source along a radial axis. An important uniformity improvement was observed for the reconstructed image of the hot spot phantom when depth of interaction correction was applied. A moderate improvement of the spatial resolution was observed when reconstructing the images of the point source with depth of interaction correction.

Approximate k-NN delta test minimization method using genetic algorithms: Application to time series

In many real world problems, the existence of irrelevant input variables (features) hinders the predictive quality of the models used to estimate the output variables. In particular, time series prediction often involves building large regressors of artificial variables that can contain irrelevant or misleading information. Many techniques have arisen to confront the problem of accurate variable selection, including both local and global search strategies. This paper presents a method based on genetic algorithms that intends to find a global optimum set of input variables that minimize the Delta Test criterion. The execution speed has been enhanced by substituting the exact nearest neighbor computation by its approximate version. The problems of scaling and projection of variables have been addressed. The developed method works in conjunction with MATLABs Genetic Algorithm and Direct Search Toolbox. The goodness of the proposed methodology has been evaluated on several popular time series examples, and also generalized to other non-time-series datasets.

Design and Calibration of a Small Animal Pet Scanner Based on Continuous LYSO Crystals and PSPMTs

We report on the design of a small animal PET scanner based on continuous LYSO crystals and position sensitive photomultiplier tubes (PSPMTs), together with the first results from the calibration. The scanner consists of eight compact modules forming an octagon and leaving a port of 110 mm aperture. Each module is made out of a continuous LYSO crystal and a PSPMT, and contains its associated electronics together with its power supply. For each module, five signals are read, summarizing all the information coming out from its 64 anode pads. Therefore, for the whole scanner only 40 signals are digitized. A calibration procedure has been implemented, measuring a spatial resolution of approximately 1.5 mm at the center of the field of view and an energy resolution of 18%. The sensitivity of the system at the center of the field of view, using only 4 modules, is observed to be of about 1%.

FPGA Custom DSP for ECG Signal Analysis and Compression

This work describes a Virtex-II implementation of a custom DSP for QRS-Complex detection, ECG signal analysis and data compression for optimum transmission and storage. For QRS-Complex detection we introduce a custom architecture based on a modification of the Hamilton-Tompkins (HT) algorithm oriented to area saving. We also use biorthogonal wavelet transform for ECG signal compression and main ECG parameters estimation. In contrast with previously published works, our modified version of the HT algorithm offers best performance (a 99% of QRS-detection over normalized noisy ECGs). Moreover, a compression ratio of 20:1 is obtained when the wavelet-based engine is running. Results have been successfully verified by using a combination of MATLAB with SystemGen, Modelsim and a FPGA PCI-based Card (AlphaData ADM-XRC-II). The QRS-complex detector and compressor requires minimum area resources in term of LUT and registers, allowing a custom DSP as coprocessor in a SoC for biomedical applications.

Faster and exact implementation of the continuous cellular automaton for anisotropic etching simulations

The current success of the continuous cellular automata for the simulation of anisotropic wet chemical etching of silicon in microengineering applications is based on a relatively fast, approximate, constant time stepping implementation (CTS), whose accuracy against the exact algorithm?a computationally slow, variable time stepping implementation (VTS)?has not been previously analyzed in detail. In this study we show that the CTS implementation can generate moderately wrong etch rates and overall etching fronts, thus justifying the presentation of a novel, exact reformulation of the VTS implementation based on a new state variable, referred to as the predicted removal time (PRT), and the use of a self-balanced binary search tree that enables storage and efficient access to the PRT values in each time step in order to quickly remove the corresponding surface atom/s. The proposed PRT method reduces the simulation cost of the exact implementation from to without introducing any model simplifications. This enables more precise simulations (only limited by numerical precision errors) with affordable computational times that are similar to the less precise CTS implementation and even faster for low reactivity systems.