T. Tollenaere

Original Contribution: A modular artificial neural network for texture processing

This paper presents a new network-based model for segregating broadband noise textures. The model starts with the oriented local energy maps obtained from filtering the textures with a bank of quadrature pair Gabor filters with different preferred orientations and spatial frequencies, and squaring and summing the quadrature pair filter outputs point-wise. Rather than detecting differences in first-order statistics from these maps, a sequence of two network modules is used for each spatial frequency channel. The modules are based on the Entropy Driven Artificial Neural Network (EDANN) model, a previously developed adaptive network module for line- and edge detection. The first EDANN module performs orientation extraction and the second performs filling-in of missing orientation information. The aim of both network modules is to produce a reliable texture segregation based on an enlarged local difference in first-order statistics in the mean and at the same time a reduced importance of differences in spatial variability; the texture boundary is detected using a third EDANN module, following the second one. Other major features of the model are: (a) texture segregation proceeds in each spatial frequency/orientation channel separately, and (b) texture segregation as well as texture boundary detection can be performed using the same core network module.

Thrombolytic profiles of clot-targeted plasminogen activators. Parameters determining potency and initial and maximal rates.

BackgroundTargeting of plasminogen activators to the thrombus by means of fibrin-specific monoclonal antibodies may enhance their thrombolytic potency. The kinetics of clot binding of two humanfibrin-specific monoclonal antibodies (MA-12B3 and MA-15C5) and of clot lysiswith their chemical 1:1 stoichiometric complexes with recombinant single-chain urokinase-type plasminogen activator (rscu-PA)(rscu-PA/MA-12B3 and rscu-PA/MA-15C5) were determined in hamsters and rabbits. Thrombolyticpotencies, maximal rates of clot lysis, and the duration of the lag phases before clot lysis of theantibody/rscu-PA conjugates were compared with those of rscu-PA and tissue-type plasminogen activator (rt-PA). Methods and ResultsBolus injection of 7.5 μg of 125I-labeled antibody in rabbits with an extracorporeal arteriovenous loopcontaining a 0.3-mL human plasma clot produced clot-to-blood ratios of 6.6±1.0 (mean±SEM) for MA-12B3 and 1.1±0.15 for MA-15C5 (p<0.001 versus MA-12B3) within 6 hours. Progressive digestion of the clot did not alter the binding of MA-12B3 but resulted in as much as a 10-fold increase of the binding of MA-15C5. The conjugates infused intravenously over 90 minutes in hamsters with a human plasma clot in thepulmonary artery produced dose-related in vivo clot lysis. Thrombolytic potencies (maximal slope of the percent lysis versus dose in milligrams of u-PA equivalent per kilogram body weight) were 2,500±440 for rscu-PA/MA-12B3, 3,600±640 for rscu-PA/MA-15C5 (p=NS vs. rscu-PA/MA-12B3), 60±8 for rscu-PA (p<0.001 versus both conjugates), and 380+66 for rt-PA (p<0.001 versus both conjugates). The plasma clearances of the conjugates were fourfold to sixfold slower than those of rscu-PA and rt-PA. Maximal rates of clot lysis, determined by continuous external radioisotope scanning over the thorax, were 0.90±0.13%, 0.91+0.17%, 0.84±0.12%, and 1.1+0.16% lysis per minute for rscu-PA/MA-12B3, rscu-PA/MA-1SC5, rscu-PA, and rt-PA, respectively; these maximal rates were obtained with 0.016,0.016, 1.0, and 0.25 mg/kg, respectively, and were associated with minimal lag phases of 18±3.2, 28±4.9, 34+3.7, and 25±3.9 minutes, respectively. ConclusionThe thrombolytic potency of the rscu-PA/antifibrin conjugates is determined by their clearance, as well as by rate and extent of initial binding to clots and by changes in binding during clotlysis. Clot targeting of rscu-PA with fibrin-specific antibodies increases its thrombolytic potency but does not alter the maximal rate or the minimal lag phase of clot lysis. These parameters appear to be independent of the nature of the plasminogen activator and of targeting.

Paper: Simulating modular neural networks on message-passing multiprocessors

While simulation programs for single neural networks, running on parallel machines, always use a fixed problem decomposition and mapping strategy, we show that this is not possible for modular neural networks. We demonstrate this by analysing decomposition and mapping issues for a particular modular neural network model: the entropy-driven artificial neural network. The classic approach to simulations consists of two steps: first a data structure is built, describing the problem to be simulated. For a neural network this data structure contains the network topology, the interconnection strengths, etc. In a second step, this data structure is read into the simulation program, which performs a fixed decomposition and mapping before simulation can take place. Since this approach cannot be used any more for simulations of modular networks, we propose a new, three-step approach, in which decomposition and mapping are taken out of the simulation program. A compiler is used to prepare the problem data structure, a splitter program takes care of problem decomposition, and the simulator program takes the decomposed problem as its input. Since all decisions with respect to decomposition and mapping are taken by the splitter, the simulator program is independent of decomposition and mapping, and hence it can handle any decomposition and mapping. Following this approach, a machine-independent simulation environment was designed, and this design was implemented on a transputer system. To show that our approach is generic (i.e. not limited to simulations of modular networks) an implementation of a Hopfield network for image restoration is described. In spite of the classic preconceptions about generic software, performance analysis and benchmark results show that our novel, generic approach can be implemented efficiently on transputer arrays.

Parallel implementation and capabilities of entropy-driven artificial neural networks

Abstract It is shown how the recently introduced Entropy-Driven Artificial Neural Network Model (EDANN) can be implemented on a parallel transputer array, using a simulation environment which makes all decomposition and mapping issues transparent. Then, using the parallel simulator, the EDANNs capabilities are exemplified in the case of orientation extraction from retinal images. By means of simulations on the parallel machine, it is shown that the EDANN is able to adapt itself optimally to the stimulus it receives, and that the same network topology is able to accomplish both 1-D and 2-D orientation inference tasks.

DECOMPOSITION AND MAPPING OF LOCALLY CONNECTED LAYERED NEURAL NETWORKS ON MESSAGE-PASSING MULTIPROCESSORS

In this paper we present an integrated model for decomposition and mapping (D&M) of Locally Connected Layered Neural Networks (LCLNs) on message-passing multiprocessors. Within the framework of this model we analyze two previously proposed D&M strategies for a particular class of LCLNs. The model is compared with the performance of a neural network simulation environment, running on a transputer array. We find that both strategies may be applicable, depending on the network size, and we can determine the problem size at which one strategy is preferred over the other. Furthermore, we find that the regularity of the communication pattern between the processors is an unexpected factor in the D&M decision.

Line-end detection and boundary gap completion in an EDANN module for orientation

Explores two sources of inaccuracies originating from the use of local line detectors for inferring curve and boundary traces: (1) due to the position uncertainty of the local line detectors, ends of thin lines are not easily detected, even if cross-orientation inhibition is applied; and (2) due to the limited ability of the local line detectors to assess more global trace information gaps appear in the curve and boundary extracted. It is shown how a single EDANN (entropy drive artificial neural networks) module processing the orientation of illumination contrast compensates for these inaccuracies by performing a two-stage detection process, a competitive and a cooperative one. In the competitive stage, a vector field of tangents to curves and boundaries is extracted by using elongated receptive fields. In the cooperative stage, line-ends are extracted and boundary gaps are bridged by broadening the neurons orientation tuning curves.<<ETX>>

Distinguishing line detection from texture segregation using a modular network-based model

An important early vision problem on how a bank of local spatial filters can be common to both line- and edge detection, and texture segregation is discussed. The authors introduce a network-based model for line- and edge detection and texture segregation. The network is based on the entropy driven artificial neural network (EDANN) model, a previously developed network module. Using a hierarchy of different instantiations of the same EDANN module, the authors were able not only to resolve the major ambiguities with line- and edge detection and texture segregation, but also to distinguish these tasks and to discount for the effect of the illuminant without relying on a diffusive filling-in process.<<ETX>>

An adaptive neural network model for distinguishing line- and edge detection from texture segregation

The authors consider an important paradigm in vision: distinguishing object contours or edges (and lines) from object surface textures. To accomplish this, an artificial neural network model, called the EDANN model, is used for both texture segregation and line and edge detection starting from a common bank of spatial filters. The model provides different representations of a retinal image in such a way that different actions and decisions about the presence of objects in the visual scene can be undertaken in a further stage. Three possible cases of distinguishing luminance-defined lines and edges from noise textures are considered.<<ETX>>

Massive MIMD neural network simulations: The connections dilemma

We present two strategies for the simulation of massive neural networks on message-passing MIMD machines. In the first strategy all interconnections between neurons are stored explicitly in interconnection matrices. During simulation, every processor is responsible for certain submatrices of these interconnection matrices. The fact that message-passing MIMD processors do not provide virtual memory seriously limits the size of the networks that can be simulated, since interconnection matrices require huge amounts of memory. An alternative strategy is not to store the connections explicitly, but generate the interconnections as they are needed. This circumvents memory limitations, but because interconnections need to be generated multiple times, it is inherently slower than the first implementation. This yields the connections dilemma: the choice between fast simulation of small networks as against slower simulation of massive networks. We present, analyze and bench-mark parallel implementations for both strategies. An efficient connection-look-up algorithm, which can be used for any network with static interconnections, ensures that simulation times for the second strategy are only marginally longer than for the first strategy. We show that for our users the connections dilemma is no longer a dilemma: by means of our look-up algorithm the simulation of massive networks becomes possible; furthermore the time to design and construct a network, prior to simulation, is considerably shorter than it is for the matrix version, and in addition this time is independent of network size. Although we have implemented both strategies on a parallel computer, the algorithms presented here can be used on any machine with memory limitations, such as personal computers.

Efficient Simulation of Massive Neural Networks on Machines with Limited Memory

This paper describes techniques for the implementation of simulation programs for large neural networks, on machines without virtual memory, such as personal computers and MIMD multiprocessors like transputer arrays and hypercubes.