This paper presents an investigation of odor localization by groups of autonomous mobile robots using principles of swarm intelligence. We describe a distributed algorithm by which groups of agents can solve the full odor localization task more efficiently than a single agent. We then demonstrate that a group of real robots under fully distributed control can successfully traverse a real odor plume. Finally, we show that an embodied simulator can faithfully reproduce the real robots experiments and thus can be a useful tool for off-line study and optimization of odor localization in the real world.

This paper presents a scalable threshold-based algorithm for allocating workers to a given task whose demand evolves dynamically over time. The algorithm is fully distributed and solely based on the local perceptions of the individuals. Each agent decides autonomously and deterministically to work only when it "feels" that some work needs to be done based on its sensory inputs. In this paper, we applied the worker allocation algorithm to a collective manipulation case study concerned with the gathering and clustering of initially scattered small objects. The aggregation experiment has been studied at three different experimental levels by using macroscopic and microscopic probabilistic models, and embodied simulations. Results show that teams using a number of active workers dynamically controlled by the allocation algorithm achieve similar or better performances in aggregation than those characterized by a constant team size, while using a considerably reduced number of agents over the whole aggregation process. Since this algorithm does not imply any form of explicit communication among agents, it represents a cost-effective solution for controlling the number of active workers in embedded systems consisting of a few to thousands of units.

We present an integrated chemical sensor array fabricated using a CMOS process followed by post-processing. The sensor presented in this paper incorporates 324 individually addressable sensing nodes. Post processing involves an electroless nickel and gold plating step to fabricate sensing contacts, and the deposition of a carbon black based polymer sensor material. The operation of the integrated sensor is confirmed. This sensor technology will allow the creation of large arrays of chemically diverse sensors.

In this paper the application of neural networks to some of the network management tasks carried out in a regional Bell telephone company is described. Network managers monitor the telephone network for abnormal conditions and have the ability to place controls in the network to improve traffic flow. Conclusions are drawn regarding the utility and effectiveness of the neural networks in automating the network management tasks.

Many robotics and navigation systems utilizing stereopsis to determine depth have rigid size and power constraints and require direct physical implementation of the stereo algorithm. The main challenges lie in managing the communication between image sensor and image processor arrays, and in parallelizing the computation to determine stereo correspondence between image pixels in real-time. This paper describes the first comprehensive system level demonstration of a dedicated low-power analog VLSI (very large scale integration) architecture for stereo correspondence suitable for real-time implementation. The inputs to the implemented chip are the ordered pixels from a stereo image pair, and the output is a two-dimensional disparity map. The approach combines biologically inspired silicon modeling with the necessary interfacing options for a complete practical solution that can be built with currently available technology in a compact package. Furthermore, the strategy employed considers multiple factors that may degrade performance, including the spatial correlations in images and the inherent accuracy limitations of analog hardware, and augments the design with countermeasures.

A given (overcomplete) discrete oriented pyramid may be converted into a steerable pyramid by interpolation. We present a technique for deriving the optimal interpolation functions (otherwise called 'steering coefficients'). The proposed scheme is demonstrated on a computationally efficient oriented pyramid, which is a variation on the Burt and Adelson (1983) pyramid. We apply the generated steerable pyramid to orientation-invariant texture analysis in order to demonstrate its excellent rotational isotropy. High classification rates and precise rotation identification are demonstrated.>

A rotation-invariant texture recognition system is presented. A steerable oriented pyramid is used to extract representative features for the input textures. The steerability of the filter set allows a shift to an invariant representation via a DFT-encoding step. Supervised classification follows. State-of-the-art recognition results are presented on a 30 texture database with a comparison across the performance of the k-NN, backpropagation and rule-based classifiers. In addition, high accuracy estimation of the input rotation angle is demonstrated.

We describe a texture analysis system in which informative discrimination rules are learned from a multiresolution representation of time textured input. The system incorporates unsupervised and supervised learning via statistical machine learning and rule-based neural networks, respectively. The textured input is represented in the frequency-orientation space via a log-Gabor pyramidal decomposition. In the unsupervised learning stage a statistical clustering scheme is used for the quantization of the feature-vector attributes. A supervised stage follows in which labeling of the textured map is achieved using a rule-based network. Simulation results for the texture classification task are given. An application of the system to real-world problems is demonstrated.>

The authors present a method for learning fuzzy logic membership functions and rules to approximate a numerical function from a set of examples of the function's independent variables and the resulting function value. This method uses a three-step approach to building a complete function approximation system: first, learning the membership functions and creating a cell-based rule representation; second, simplifying the cell-based rules using an information-theoretic approach for induction of rules from discrete-valued data; and, finally, constructing a computational (neural) network to compute the function value given its independent variables. This function approximation system is demonstrated with a simple control example: learning the truck and trailer backer-upper control system.>

Describes a novel neural architecture for learning deterministic context-free grammars, or equivalently, deterministic pushdown automata. The unique feature of the proposed network is that it forms stable state representations during learning-previous work has shown that conventional analog recurrent networks can be inherently unstable in that they cannot retain their state memory for long input strings. The authors have previously introduced the discrete recurrent network architecture for learning finite-state automata. Here they extend this model to include a discrete external stack with discrete symbols. A composite error function is described to handle the different situations encountered in learning. The pseudo-gradient learning method (introduced in previous work) is in turn extended for the minimization of these error functions. Empirical trials validating the effectiveness of the pseudo-gradient learning method are presented, for networks both with and without an external stack. Experimental results show that the new networks are successful in learning some simple pushdown automata, though overfitting and non-convergent learning can also occur. Once learned, the internal representation of the network is provably stable; i.e., it classifies unseen strings of arbitrary length with 100% accuracy.

Recent work has shown that recurrent neural networks have the ability to learn finite state automata from examples. In particular, networks using second-order units have been successful at this task. In studying the performance and learning behavior of such networks we have found that the second-order network model attempts to form clusters in activation space as its internal representation of states. However, these learned states become unstable as longer and longer test input strings are presented to the network. In essence, the network “forgets” where the individual states are in activation space. In this paper we propose a new method to force such a network to learn stable states by introducing discretization into the network and using a pseudo-gradient learning rule to perform training. The essence of the learning rule is that in doing gradient descent, it makes use of the gradient of a sigmoid function as a heuristic hint in place of that of the hard-limiting function, while still using the discretized value in the feedback update path. The new structure uses isolated points in activation space instead of vague clusters as its internal representation of states. It is shown to have similar capabilities in learning finite state automata as the original network, but without the instability problem. The proposed pseudo-gradient learning rule may also be used as a basis for training other types of networks that have hard-limiting threshold activation functions.

Various aspects of single-phased burst-error-correcting array codes are explored. These codes are composed of two-dimensional arrays with row and column parities with a diagonally cyclic readout order; they are capable of correcting a single burst error along one diagonal. Optimal codeword sizes are found to have dimensions n/sub 1/*n/sub 2/ such that n/sub 2/ is the smallest prime number larger than n/sub 1/. These codes are capable of reaching the Singleton bound. A new type of error, approximate errors, is defined; in q-ary applications, these errors cause data to be slightly corrupted and therefore still close to the true data level. Phased burst array codes can be tailored to correct these codes with even higher rates than before.>

A loss function, or objective function, is a function used to compare parameters when fitting a model to data. The loss function gives a distance between the model output and the desired output. Two common examples are the squared-error loss function and the cross entropy loss function. Minimizing the mean-square error loss function is equivalent to minimizing the mean square difference between the model output and the expected value of the output given a particular input. This property of minimization to the expected value is formalized as P-admissibility. The necessary and sufficient conditions for P-admissibility, leading to a parametric description of all P-admissible loss functions, are found. In particular, it is shown that two of the simplest members of this class of functions are the squared error and the cross entropy loss functions. One application of this work is in the choice of a loss function for training neural networks to provide probability estimates.>

In this paper we propose a network architecture that combines a rule-based approach with that of the neural network paradigm. Our primary motivation for this is to ensure that the knowledge embodied in the network is explicitly encoded in the form of understandable rules. This enables the network's decision to be understood, and provides an audit trail of how that decision was arrived at. We utilize an information theoretic approach to learning a model of the domain knowledge from examples. This model takes the form of a set of probabilistic conjunctive rules between discrete input evidence variables and output class variables. These rules are then mapped onto the weights and nodes of a feedforward neural network resulting in a directly specified architecture. The network acts as parallel Bayesian classifier, but more importantly, can also output posterior probability estimates of the class variables. Empirical tests on a number of data sets show that the rule-based classifier performs comparably with standard neural network classifiers, while possessing unique advantages in terms of knowledge representation and probability estimation.

An algorithm for the induction of rules from examples is introduced. The algorithm is novel in the sense that it not only learns rules for a given concept (classification), but it simultaneously learns rules relating multiple concepts. This type of learning, known as generalized rule induction, is considerably more general than existing algorithms, which tend to be classification oriented. Initially, it is focused on the problem of determining a quantitative, well-defined rule preference measure. In particular, a quantity called the J-measure is proposed as an information-theoretic alternative to existing approaches. The J-measure quantifies the information content of a rule or a hypothesis. The information theoretic origins of this measure are outlined, and its plausibility as a hypothesis preference measure is examined. The ITRULE algorithm, which uses the measure to learn a set of optimal rules from a set of data samples, is defined. Experimental results on real-world data are analyzed.>

The mean lifetimes are studied of semiconductor memories that have been encoded with an on-chip single error-correcting code along each row of memory cells. Specifically, the effects of single-cell soft errors and various hardware failures (single-cell, row, column, row-column, and entire chip) in the presence of soft-error scrubbing are examined. An expression is presented for computing the mean time to failure of such memories in the presence of these types of errors using the Poisson approximation; the expression has been confirmed experimentally to accurately model the mean time to failure of memories protected by single error-correcting codes. These analyses will enable the system designer to accurately assess the improvement in mean time to failure (MTTF) achieved by the use of error-control coding.>

A model for a class of high-capacity associative memories is presented. Since they are based on two-layer recurrent neural networks and their operations depend on the correlation measure, these associative memories are called recurrent correlation associative memories (RCAMs). The RCAMs are shown to be asymptotically stable in both synchronous and asynchronous (sequential) update modes as long as their weighting functions are continuous and monotone nondecreasing. In particular, a high-capacity RCAM named the exponential correlation associative memory (ECAM) is proposed. The asymptotic storage capacity of the ECAM scales exponentially with the length of memory patterns, and it meets the ultimate upper bound for the capacity of associative memories. The asymptotic storage capacity of the ECAM with limited dynamic range in its exponentiation nodes is found to be proportional to that dynamic range. Design and fabrication of a 3-mm CMOS ECAM chip is reported. The prototype chip can store 32 24-bit memory patterns, and its speed is higher than one associative recall operation every 3 mus. An application of the ECAM chip to vector quantization is also described.

A neural network approach to the problem of color constancy is presented. Various algorithms based on Land's retinex theory are discussed with respect to neurobiological parallels, computational efficiency, and suitability for VLSI implementation. The efficiency of one algorithm is improved by the application of resistive grids and is tested in computer simulations; the simulations make clear the strengths and weaknesses of the algorithm. A novel extension to the algorithm is developed to address its weaknesses. An electronic system that is based on the original algorithm and that operates at video rates was built using subthreshold analog CMOS VLSI resistive grids. The system displays color constancy abilities and qualitatively mimics aspects of human color perception.

Information set decoding is an algorithm for decoding any linear code. Expressions for the complexity of the procedure that are logarithmically exact for virtually all codes are presented. The expressions cover the cases of complete minimum distance decoding and bounded hard-decision decoding, as well as the important case of bounded soft-decision decoding. It is demonstrated that these results are vastly better than those for the trivial algorithms of searching through all codewords or through all syndromes, and are significantly better than those for any other general algorithm currently known. For codes over large symbol fields, the procedure tends towards a complexity that is subexponential in the symbol size.>

A central paradox of coding theory concerns the existence and construction of the best codes. Virtually every linear code is good, in the sense that it meets the Gilbert-Varshamov bound on distance versus redundancy. Despite the sophisticated constructions for codes derived over the years, however, no one has succeeded in demonstrating a constructive procedure that yields such codes over arbitrary symbol fields. Using the theory of Kolmogorov complexity, it is shown that this statement holds true in a rigorous mathematical sense: any linear code that is truly random, in the sense that there is no concise way of specifying the code, is good. Furthermore, random selection of a code that does contain some constructive pattern results, with probability bounded away from zero, in a code that does not meet the Gilbert-Varshamov bound regardless of the block length of the code. In contrast to the situation for linear codes, it is shown that there are effectively random nonlinear codes which have no guarantee on distance. In addition, it is shown that the techniques of Kolmogorov complexity can be used to derive typical properties of classes of codes in a novel way.>

The lifetimes of computer memories which are protected with single-error-correcting-double-error-detecting (SEC-DED) codes are studies. The authors assume that there are five possible types of memory chip failure (single-cell, row, column, row-column and whole chip), and, after making a simplifying assumption (the Poisson assumption), have substantiated that experimentally. A simple closed-form expression is derived for the system reliability function. Using this formula and chip reliability data taken from published tables, it is possible to compute the mean time to failure for realistic memory systems.>

Linear sum codes (LSCs) form a class of error control codes designed to provide on-chip error correction to semiconductor random access memories (RAMs). They use the natural addressing scheme found on RAMs to form and access codewords with a minimum of overhead. The authors formally define linear sum codes and examine some of their characteristics. Specifically, they examine their minimum distance characteristics, their error correcting capabilities, and the complexity involved in their implementation. In addition, detailed consideration is given to an easily implemented class of single-, double-, and triple-error correcting LSCs.>

A communication theory approach to decision tree design based on a top-town mutual information algorithm is presented. It is shown that this algorithm is equivalent to a form of Shannon-Fano prefix coding, and several fundamental bounds relating decision-tree parameters are derived. The bounds are used in conjunction with a rate-distortion interpretation of tree design to explain several phenomena previously observed in practical decision-tree design. A termination rule for the algorithm called the delta-entropy rule is proposed that improves its robustness in the presence of noise. Simulation results are presented, showing that the tree classifiers derived by the algorithm compare favourably to the single nearest neighbour classifier.>