This paper presents a new semi-supervised neural architecture that learns to classify objects at a distance through experience. It utilizes Fuzzy LAPART extended with two temporal integrator subnetworks to create time-stamped perceptual memory codes in an unsupervised manner during object approach, and to retrospectively learn class code inferences at contact. Fuzzy LAPART, Temporal Integrators and the integrated architecture are presented. Next, the agent-based modeling and neural simulator tools used to model this architecture are described. Finally, a study is presented that illustrates the learning performance of this architecture embodied in a simple simulated agent moving in a 2D environment.

Hippocampus within medial temporal lobe of the brain is essentially involved in episodic memory formation. Rather than simply being a mechanism of storing information, episodic memory associates information such as the spatial and temporal context of an event. Using hippocampus neurophysiology and functionality as an inspiration, we have developed an artificial neural network architecture called Associative-ART to associate k-tuples of inputs. In this paper we present an overview of hippocampus neurophysiology, explain the design of our neural network architecture, and present experimental results from an implementation of our architecture.

A category-theoretic account of neural network semantics has been used to characterize concept representation in neural memory. This involves categories of objects and morphisms representing the activity in connectionist structures at different stages of weight adaptation. The definition used in this previous work for the notion of a neural morphism does not allow temporal memories to be retrieved stepwise as event sequences-all events must be retrieved simultaneously. An extended definition is proposed that enables episodic information to be retrieved in time sequence.

Brain connections formed during the nurturing period of an infant's development are fundamental for survival. In this paper, elementary brain (neural interconnection pattern) evolution is simulated for various individuals in two similar artificial species. The simulation yields information about the learning, performance and brain structure of the population over time. Concepts from categorical neural semantic theory (CNST) are used to analyze the development of neural structure as evolution progresses. FlatWorld, a virtual two dimensional environment, is used to test survival skills of simple embodied neural agents. A combination of genetic algorithms (GA) and neural networks (NN) is applied within FlatWorld to study the relationship between the nurturing of the infant individuals during their developmental period with their subsequent behavior in the environment and the evolution of the associated brain structures. The results show evidence that during evolution, learning performance increases when brain structures required from CNST are formed, and that survival skills increase over evolutionary time-scales due to the formation of these structures.

Encoding sensor observations across time is a critical component in the ability to model cognitive processes. All biological cognitive systems receive sensory stimuli as continuous streams of observed data over time. Therefore, the perceptual grounding of all biological cognitive processing is in temporal semantic encodings, where the particular grounding semantics are sensor modalities. We introduce a technique that encodes temporal semantic data as temporally integrated patterns stored in adaptive resonance theory (ART) modules.

A recent neural network semantic theory provides the framework for mapping ontologies to neural networks. We use category theory, the mathematical theory of structure, to explore the concept representational abilities of select neural networks. Methodologies suggested by the semantic theory have been gainfully applied to specific applications. This paper describes a rigorous and numerical study of the implementation of neural category representations into an actual neural network.

The use of neural network architectures has historically presented a challenge to engineers. Problem domains could be "learned", but the acquired knowledge could be extracted only under limited circumstances. Healy and Caudell's application of category theory has been shown to improve both architecture design and performance. This paper reports on the application of category theory to the design of a simulated robot control system, where the neural network controller is constructed based upon a desired conceptual ontology. Three experiments then explore the implications of this approach on the prediction and improvement of robot performance.

While learning algorithms have been used for astronomical data analysis, the vast majority of those algorithms have used supervised learning. In a continuation of the work described in Young et ah [18] we examine the use of unsupervised learning for this task with two types of Adaptive Resonance Theory (ART) neural networks. Using synthetic astronomical data from SkyMaker[2], [3] which was designed to mimic the dynamic range of the CTI-[14] telescope, we compared the ability of the ART-1 neural network[4] and the ART-1 neural network with category theoretic modiflcation[9], [11] to detect regions of interest and to characterize noise. We show a difference in the geometries of the templates created by each architecture. We also show an analysis of the two architectures over a range of parameter settings. The results provided show that ART neural networks and unsupervised learning algorithms in general should not be overlooked for astronomical data analysis.

Concepts have been expressed mathematically as propositions in a distributive lattice. A more comprehensive formulation is that of a generalized lattice, or category, in which the concepts are related in hierarchical fashion by lattice-like links called concept morphisms. A concept morphism describes how a more abstract concept is used within a more specialized concept, as the color "red" is used in describing "apples". Often, an abstract concept can be used in a more specialized concept in more than one way as with "color", which can appear in "apples" as either "red", "yellow" or "green". Further, "color" appears in "apples" because it appears in "red", "yellow" and "green", which in turn appear in "apples", expressed via the composition of concept morphisms. Using categorical constructs based upon composition together with structure-preserving mappings that preserve compositional structure, a recently-developed semantic theory shows how abstract and specialized concepts are learned by a neural network.

The semantics of neural networks can be analyzed mathematically as a distributed system of knowledge and as systems of possible worlds expressed in the knowledge. Learning in a neural network can be analyzed as an attempt to acquire a representation of knowledge. We express the knowledge system, systems of possible worlds, and neural architectures at different stages of learning as categories. Diagrammatic constructs express learning in terms of pre-existing knowledge representations. Functors express structure-preserving associations between the categories. This analysis provides a mathematical vehicle for understanding connectionist systems and yields design principles for advancing the state of the art.

Many studies have addressed the knowledge representation capability of neural networks. A recently-developed mathematical semantic theory explains the relationship between knowledge and its representation in connectionist systems. The theory yields design principles for neural networks whose behavioral repertoire expresses any desired capability that can be expressed logically. In this paper, we show how the design principle of limit formation can he applied to modify the ART-1 architecture, yielding a discrimination capability that goes beyond vigilance. Simulations of this new design illustrate the increased discrimination ability it provides for multi-spectral image analysis.

Our analytics task is to identify, characterize, and visualize anomalous subsets of as large of a collection of network connection data as possible. We use a combination of HPC resources, advanced algorithms, and visualization techniques. To effectively and efficiently identify the salient portions of the data, we rely on a multistage workflow that includes data acquisition, summarization (feature extraction), novelty detection, and classification. Once these subsets of interest have been identified and automatically characterized, we use a stateof- the-art high-dimensional query system to extract this data for interactive visualization. Our approach is equally useful for other large-data analysis problems where it is more practical to identify interesting subsets of the data for visualization than it is to render all data elements. By reducing the size of the rendering workload, we enable highly interactive and useful visualizations.

This paper describes a general methodological framework for evaluating the perceptual properties of auditory stimuli. The framework provides analysis techniques that can ensure the effective use of sound for a variety of applications, including virtual reality and data sonification systems. Specifically, we discuss data collection techniques for the perceptual qualities of single auditory stimuli including identification tasks, context-based ratings, and attribute ratings. In addition, we present methods for comparing auditory stimuli, such as discrimination tasks, similarity ratings, and sorting tasks. Finally, we discuss statistical techniques that focus on the perceptual relations among stimuli, such as Multidimensional Scaling (MDS) and Pathfinder Analysis. These methods are presented as a starting point for an organized and systematic approach for nonexperts in perceptual experimental methods, rather than as a complete manual for performing the statistical techniques and data collection methods. It is our hope that this paper will help foster further interdisciplinary collaboration among perceptual researchers, designers, engineers, and others in the development of effective auditory displays.

Stochastic, or nonpitched, sounds fill our real-world environment. Humans almost continuously hear stochastic sounds, such as wind, rain, motor sounds, and different types of impact sounds. Because of their prevalence in real-world environments, it is important to include these types of sounds for realistic virtual environment simulations. This paper describes a synthesis approach that uses wavelets for modeling stochastic-based sounds. Parameterizations of the wavelet models yield a variety of related sounds from a small set of models. The result is dynamic sound models that can change according to changes in the virtual environment. This paper contains a description of the sound synthesis process, several developed models, and the on-going perceptual experiments for validating the sound synthesis veracity. The developed models and results demonstrate proof of the concept and illustrate the potential of this approach.

Sound is a critical component of multimodal representations of information in virtual environments. Our research program continues to develop and evaluate sound software system architectures, localization techniques, sound synthesis, and musical sonification as tools to enhance human comprehension of complex data sets.

We introduce a new architecture designed by applying a recently-developed mathematical model of neural network semantics using category theory. The new design has multiple subnetworks associated with different sensors and association regions. The subnetworks form individual, hierarchical representations of a body of knowledge. Subnetwork interconnections adapt to link the individual concept representations appropriately and provide knowledge coherence, representing a single knowledge hierarchy across the multi-sensor network.

Visualization is a useful method for understanding both learning and computation in artificial neural networks. There are a large number of parameters in a neural network. By viewing these parameters pictorially, a better understanding can be gained of how a network maps inputs to outputs. eLoom is an open source graph simulation tool, developed at the University of New Mexico, that enables users to specify and simulate various neural network models. Its specification language enables users to construct and simulate arbitrary, potentially hierarchical network structures on serial and parallel processing systems. In addition, eLoom is integrated with UNM's Flatland, an open source virtual environment development tool to provide real-time visualizations of the network structure and activity. ART-1 and LAPART-II neural networks are presented to illustrate eLoom and Flatland's capabilities.

Using a neural network as an abstract black box makes it hard to grasp its inner workings. Visualizing a dynamic functioning neural network along with its related model simulation may lead to a deeper comprehension of both. We propose using a virtual environment as a tool to investigate the complex space of a neural network. As an example, we train a simulated remote autonomous vehicle to navigate a pre-planned path through a set of obstacles using a LAPART neural network. Within the virtual environment, we can easily and naturally position ourselves to best observe the activity in which we are most interested and discover the evolving space of an operating neural network.

A simplified model of the fly's early visual processing system was simulated. The model contains coupled nonlinear differential equations, increasing the complexity of software and its computational burden. An extension of the model to large arrays of retinal cartridges required high performance computing which further increased the complexity. Conventional visualization methods limited software development and model analysis. As a solution, more sophisticated visualization tools, such as IBM's open visualization data explorer and UNM's Flatland, were employed. These tools expedited software debugging and development. The accuracy and added insight afforded by these visualization tools provided a more natural environment for simulation analysis.

This paper presents the results of a computer study of supervised learning and generalization in a new neural architecture that has extremely tight bounds on learning convergence. LAPART-2 is an extended version of the LAPART-1 introduced previously by the authors (1998). This paper explores the architectural generalisation through a series of numerical experiments using challenging problems in classification. This class of problem is used in this study because of the simplicity of correctness analysis and the availability of theoretical bounds on performance. Three classification problems were picked to initially study the generalization in LAPART-2 learning. Bayesian classification performance was calculated for each of the test problems for comparison. These experiments demonstrate that the generalization performance of LAPART-2 closely matches that of Bayesian for each of the problems, and converge in two or less epochs. LAPART-2 has one of the tightest theoretical bounds on learning convergence and excellent generalization performance.

The superior colliculus (SC) is a midbrain structure containing visual, auditory, and somatosensory neuronal afferent connections. We define an artificial neural network model for use in a mobile multisensory robotic system. This model attempts to emulate SC operation with respect to its role in the generation of saccadic motor command signals sent to ocular muscle control centers in the midbrain. Much research has been centered on the exact model for this process in the visual modality. There has also been supporting research in quantifying the auditory response mapping for saccade. Our model integrates knowledge from this work along with neuroanatomical details to define a multimodal approach for understanding saccadic motor signal generation. It is important to note that we do not attempt to create an exact neurobiological equivalent function. We only use the anatomical knowledge as a motivation for our model architecture.

Using anatomy of Musca domestica's (the housefly) first optic ganglion, the lamina ganglionaris, a biologically inspired artificial neural network is presented. This paper focuses on communication phenomena and activity between specific neurons in the lamina and models nonsynaptic communication between neurons. Test results demonstrate that nonsynaptic communication between neurons can achieve results similar to synaptic connections in early vision systems.

In virtual reality (VR), head movement is tracked through inertial and optical sensors. Computation and communication times result in delays between measurements and updating of the new frame in the head mounted display(HMD). These delays result in problems, including motion sickness. We use recurrent and time delay neural networks to predict the head location and use it to calculate the new frame. A predictability analysis is used in designing the prediction system.

We present a theoretical analysis of a version of the LAPART adaptive inferencing neural network. Our main result is a proof that the new architecture, called LAPART 2, converges in two passes through a fixed training set of inputs. We also prove that it does not suffer from template proliferation. For comparison, Georgiopoulos et al. have proved the upper bound n-1 on the number of passes required for convergence for the ARTMAP architecture, where n is the size of the binary pattern input space. If the ARTMAP result is regarded as an n-pass, or finite-pass, convergence result, ours is then a two-pass, or fixed-pass, convergence result. Our results have added significance in that they apply to set-valued mappings, as opposed to the usual supervised learning model of affixing labels to classes.

Envisioning neural networks as systems that learn rules calls forth the verification issues already being studied in knowledge-based systems engineering, and complicates these with neural-network concepts such as nonlinear dynamics and distributed memories. We show that the issues can be clarified and the learned rules visualized symbolically by formalizing the semantics of rule-learning in the mathematical language of two-valued predicate logic. We further show that this can, at least in some cases, be done with a fairly simple logical model. We illustrate this with a combination of two example neural-network architectures, LAPART, designed to learn rules as logical inferences from binary data patterns, and the stack interval network, which converts real-valued data into binary patterns that preserve the semantics of the ordering of real values. We discuss the significance of the formal model in facilitating the analysis of the underlying logic of rule-learning and numerical data representation. We provide examples to illustrate the formal model, with the combined stack interval/LAPART networks extracting rules from numerical data.

We describe a neural information retrieval system (NIRS), now in production within the Boeing Company, which has been developed for the identification and retrieval of engineering designs. Two-dimensional and three-dimensional representations of engineering designs are input to adaptive resonance theory (ART-1) neural networks to produce clusters of similar parts. The trained networks are then used to recall an appropriate cluster when queried with a new part design. This application is of great practical value to industry because it aids in the identification, retrieval, and reuse of engineering designs, potentially saving large amounts of nonrecurring costs. In this paper, we review the application, the neural architectures and algorithms, and then give the current status and the lessons learned in developing a neural network system for production use in industry.

LAPART, a neural network architecture for logical inferencing and supervised learning is discussed. Emphasizing its use in recognizing familiar sequences of patterns by verifying pattern pairs inferred from prior experience. It consists of interconnected adaptive resonance theory (ART) networks. The interconnects enable LAPART to learn to infer one pattern class from another to form a predictive sequence. It predicts a next pattern class based upon recognition of a current pattern and tests the prediction as new data become available. A confirmed prediction aids verification of a familiar sequence, and a disconfirmation flags a novel pairing of patterns. A simulation of LAPART is applied to verification of a hypothetical, known target using a sequence of sensor images obtained along a predetermined approach path. Application issues are addressed with a simple strategy, and it is shown how they could be addressed in a more complete fashion. Other topics, including a logical interpretation of ART and LAPART, are discussed.

A solution to the problem of implementation of the adaptive resonance theory (ART) of neural networks that uses an optical correlator which allows the large body of correlator research to be leveraged in the implementation of ART is presented. The implementation takes advantage of the fact that one ART-based architecture, known as ART1, can be broken into several parts, some of which are better to implement in parallel. The control structure of ART, often regarded as its most complex part, is actually not very time consuming and can be done in electronics. The bottom-up and top-down gated pathways, however, are very time consuming to simulate and are difficult to implement directly in electronics due to the high number of interconnections. In addition to the design, the authors present experiments with a laboratory prototype to illustrate its feasibility and to discuss implementation details that arise in practice. This device can potentially outperform alternative implementations of ART1 by as much as two to three orders of magnitude in problems requiring especially large input fields.