Cris Koutsougeras

Complex domain backpropagation

The backpropagation algorithm is extended to complex domain backpropagation (CDBP) which can be used to train neural networks for which the inputs, weights, activation functions, and outputs are complex-valued. Previous derivations of CDBP were necessarily admitting activation functions that have singularities, which is highly undesirable. In the derivation, CDBP is derived so that that it accommodates classes of suitable activation functions. One such function is found and the circuit implementation of the corresponding neuron is given. CDBP hardware circuits can be used to process sinusoidal signals all at the same frequency (phasors). >

KAT-5: Robust Systems for Autonomous Vehicle Navigation in Challenging and Unknown Terrain

Kat-5 was the fourth vehicle to make history in DARPA’s 2005 Grand Challenge, where for the first time ever, autonomous vehicles were able to travel through 100 miles of rough terrain at average speeds greater than 15 mph. In this paper, we describe the mechanisms and methods that were used to develop the vehicle. We describe the main hardware systems with which the vehicle was outfitted for navigation, computing, and control. We describe the sensors, the computing grid, and the methods that controlled the navigation based on the sensor readings. We also discuss the experiences gained in the course of the development and provide highlights of actual field performance.

Event-driven sensor deployment using self-organizing maps

Coverage is an important optimization objective in pre and post-deployment stage of a Wireless Sensor Network (WSN). In this paper, we address the issue of placing a finite set of sensors to cover an area of arbitrary geometry. Unlike many existing works concerned with uniform coverage of a target area, we take into account the realistic consideration of the probability density for events to be sensed, termed as event-driven coverage. The objective is to distribute sensors so that the distribution density of the sensors matches that of the probability density of events to be sensed. The expected event distribution is assumed to be stationary and known a priori, directly or indirectly, in the form of sample maps. In this context we explore and evaluate the concept of Self-Organizing Maps (SOMs) and its derivative variants to address the coverage problem. Various forms of SOMs methods as well as the known methods of Virtual Fields are also compared via experimentation.

A knowledge-based expert system for automatic visual VLSI reverse-engineering: VLSI layout version

This paper presents a method of knowledge representation for very large scale integration (VLSI) chip design which provides the necessary information for abstraction from the physical design to gate-level logic through a high-level behavioral model. The representation scheme used by the ANTISTROFEAS system utilizes a hierarchical attributed graph structure which consists of incrementally abstracted design information for the VLSI system. This method of knowledge representation is well-suited to reverse-engineering of VLSI chips from the layer mask layout data, but is also applicable to applications at many levels of the design process including design rule checking, logic synthesis, design verification, and partitioning-compaction problems. The representation scheme is applicable to any VLSI technology, and is designed to take advantage of artificial intelligence. expert system techniques, by disassociating the representation and manipulation of the VLSI design data from the rules which govern its correctness and transformation for other usage.

Data flow graph partitioning to reduce communication cost

This paper presents a cost-effective scheme for partitioning large data flow graphs. Standard data flow machine architectures are assumed in this work. The objective is to reduce the overhead due to token transfers through the communication network of the machine. When this scheme is employed on large graphs, the load distribution on the rings of the data flow machine is also improved. A canonical form of a data flow graph is introduced to establish the relationship between the communication overhead and the size reduction of the partition cut-set. General lower estimates on the overhead are derived in terms of processing and transmission delay parameters of the machine. The method uses heuristics and an evaluation function to guide the partition algorithm. Some implications of the proposed method on the organization of the data flow machines are discussed.

Experiments with various neural architectures for handwritten character recognition

In this research we address the problem of recognition of isolated handwritten characters. Handwritten character recognition has been a topic of research for a long period of time. It has been argued that this problem is difficult to model using classical modeling techniques, and that neural networks offer a new perspective to approaching this problem. This paper outlines the experimental evidence we have compiled while investigating possible approaches to handwritten character recognition. It is the hypothesis of our approach that handwritten character recognition is a pattern recognition problem and that there exists a set of unique features in the data which can be used for classification.

On features used for handwritten character recognition in a neural network environment

Neural nets are considered as the underlying computing mechanism for a robust approach to the problem of handwritten character recognition. It is expected that recognition mechanisms will be developed through learning algorithms. A key factor to this problem is the set of primitive features which are used to form the raw input vectors representing the digitized image of a character. The authors have explored a number of conventional and new features that can be used in concert with adaptive clustering schemes. Experiences of the performance of these features are presented. A feature which the authors call shadow and which is presented here has produced particularly encouraging results.

SAMPLE COMPLEXITY FOR FUNCTION LEARNING TASKS THROUGH LINEAR NEURAL NETWORKS

We find new sample complexity bounds for real function learning tasks in the uniform distribution by means of linear neural networks. These bounds, tighter than the distribution-free ones reported ...

Niching in an ES/EP context

Niching or speciation is a particularly appropriate for multimodal function optimization. An irony in EC research is that genetic algorithms (GAs) are not touted primarily as function optimizers yet all reported niching research is in the GA context. Evolution strategies (ES) and major variants of evolutionary programming are better suited by design for global optimization. Borrowing methods that have been reported for niching in GAs, we have applied them to an EC algorithm that resembles ES in structure. We have found that the selection methods in ES, e.g., (/spl mu/+/spl lambda/), interact satisfactorily with niching strategies used in GAs. On the other hand, adapting selection methods such as SUS that minimize bias does not lead to favorable results. This is counter to expectations but can be reconciled with prevailing theories. We conclude with a conjecture concerning a lower bound on population size for multimodal optimization.

Handwritten character recognition using low resolutions

Abstract This paper deals with an experimental study into the recognition of handwritten characters using representations of low resolution. For this experiment several methodologies were used, such as an attributed graph OCR method, a back-propagation neural network (BNN), a recurrent neural network (RNN), and the Castor neural network (CNN). The attributed graph methodology is based on the representation (mapping) of the text characters onto a small size two-dimensional array of 12×9 cells. For the recognition process each character is considered as a composition of “main” and “secondary” features. The main features are the important parts of a character for its successful recognition. The secondary (or artistic) features are the parts of a character that contribute to its various representations. The attribute graph methodology presented in this paper attempts to prove that the recognition of a reduced-size character provides a robust approach for the recognition of handwritten text. The BNN approach is used here as a comparative recognition method, although it has some serious weaknesses at low resolution. The RNN approach for handwritten character recognition is based upon recurrent neural networks, which have a feedback mechanism. The feedback mechanism acts to integrate new values of feature vectors with their predecessors. The output is supervised according to a target function. These networks can deal with inputs and outputs that are explicit functions of time. A new way of associating shape information was used, which gives very consistent results for handwritten character recognition. In this scheme the “shadow” of each character was considered, to find the distances between the margins of the character. The distances are normalized with respect to the maximum distance in the entire shape to minimize the effect of disproportionally formed characters. In addition, the performance of Castor’s neural net was evaluated for the recognition of handwritten text characters, by using character data sets with various resolutions. For this effort the three neural networks and attributed graph approaches used a set of 5000 handwritten characters, and their results are compared.