Information Fusion and Hierarchical Knowledge Discovery by ARTMAP Neural Networks

Mapping novel terrain from sparse, complex data often requires the resolution of conflicting information from sensors working at different times, locations, and scales, and from experts with different goals and situations. Information fusion methods help resolve inconsistencies in order to distinguish correct from incorrect answers, as when evidence variously suggests that an object's class is car, truck, or airplane. The methods developed here consider a complementary problem, supposing that information from sensors and experts is reliable though inconsistent, as when evidence suggests that an objects class is car, vehicle, or man-made. Underlying relationships among objects are assumed to be unknown to the automated system of the human user. The ARTMAP information fusion system uses distributed code representations that exploit the neural network's capacity for one-to-many learning in order to produce self-organizing expert systems that discover hierarchial knowledge structures. The system infers multi-level relationships among groups of output classes, without any supervised labeling of these relationships. The procedure is illustrated with two image examples.

Self-Organizing Information Fusion and Hierarchical Knowledge Discovery: A New Framework Using Artmap Neural Networks

Classifying novel terrain or objects from sparse, complex data may require the resolution of conflicting information from sensors woring at different times, locations, and scales, and from sources with different goals and situations. Information fusion methods can help resolve inconsistencies, as when eveidence variously suggests that and object's class is car, truck, or airplane. The methods described her address a complementary problem, supposing that information from sensors and experts is reliable though inconsistent, as when evidence suggests that an object's class is car, vehicle, and man-made. Underlying relationships among classes are assumed to be unknown to the autonomated system or the human user. The ARTMAP information fusion system uses distributed code representations that exploit the neural network's capacity for one-to-many learning in order to produce self-organizing expert systems that discover hierachical knowlege structures. The fusion system infers multi-level relationships among groups of output classes, without any supervised labeling of these relationships. The procedure is illustrated with two image examples, but is not limited to image domain.

Modeling Financial Time Series with Artificial Neural Networks

Financial time series convey the decisions and actions of a population of human actors over time. Econometric and regressive models have been developed in the past decades for analyzing these time series. More recently, biologically inspired artificial neural network models have been shown to overcome some of the main challenges of traditional techniques by better exploiting the non-linear, non-stationary, and oscillatory nature of noisy, chaotic human interactions. This review paper explores the options, benefits, and weaknesses of the various forms of artificial neural networks as compared with regression techniques in the field of financial time series analysis.

On the Relationship Between the Generalized Equality Classifier and ART 2 Neural Networks

In this paper, we introduce the Generalized Equality Classifier (GEC) for use as an unsupervised clustering algorithm in categorizing analog data. GEC is based on a formal definition of inexact equality originally developed for voting in fault tolerant software applications. GEC is defined using a metric space framework. The only parameter in GEC is a scalar threshold which defines the approximate equality of two patterns. Here, we compare the characteristics of GEC to the ART2-A algorithm (Carpenter, Grossberg, and Rosen, 1991). In particular, we show that GEC with the Hamming distance performs the same optimization as ART2. Moreover, GEC has lower computational requirements than AR12 on serial machines.

Adaptive Neural Networks for Control of Movement Trajectories Invariant under Speed and Force Rescaling

This article describes two neural network modules that form part of an emerging theory of how adaptive control of goal-directed sensory-motor skills is achieved by humans and other animals. The Vector-Integration-To-Endpoint (VITE) model suggests how synchronous multi-joint trajectories are generated and performed at variable speeds. The Factorization-of-LEngth-and-TEnsion (FLETE) model suggests how outflow movement commands from a VITE model may be performed at variable force levels without a loss of positional accuracy. The invariance of positional control under speed and force rescaling sheds new light upon a familiar strategy of motor skill development: Skill learning begins with performance at low speed and low limb compliance and proceeds to higher speeds and compliances. The VITE model helps to explain many neural and behavioral data about trajectory formation, including data about neural coding within the posterior parietal cortex, motor cortex, and globus pallidus, and behavioral properties such as Woodworth's Law, Fitts Law, peak acceleration as a function of movement amplitude and duration, isotonic arm movement properties before and after arm-deafferentation, central error correction properties of isometric contractions, motor priming without overt action, velocity amplification during target switching, velocity profile invariance across different movement distances, changes in velocity profile asymmetry across different movement durations, staggered onset times for controlling linear trajectories with synchronous offset times, changes in the ratio of maximum to average velocity during discrete versus serial movements, and shared properties of arm and speech articulator movements. The FLETE model provides new insights into how spina-muscular circuits process variable forces without a loss of positional control. These results explicate the size principle of motor neuron recruitment, descending co-contractive compliance signals, Renshaw cells, Ia interneurons, fast automatic reactive control by ascending feedback from muscle spindles, slow adaptive predictive control via cerebellar learning using muscle spindle error signals to train adaptive movement gains, fractured somatotopy in the opponent organization of cerebellar learning, adaptive compensation for variable moment-arms, and force feedback from Golgi tendon organs. More generally, the models provide a computational rationale for the use of nonspecific control signals in volitional control, or "acts of will", and of efference copies and opponent processing in both reactive and adaptive motor control tasks.

The Construction of Arbitrary Stable Dynamics in Non-Linear Neural Networks

In this paper, two methods for constructing systems of ordinary differential equations realizing any fixed finite set of equilibria in any fixed finite dimension are introduced; no spurious equilibria are possible for either method. By using the first method, one can construct a system with the fewest number of equilibria, given a fixed set of attractors. Using a strict Lyapunov function for each of these differential equations, a large class of systems with the same set of equilibria is constructed. A method of fitting these nonlinear systems to trajectories is proposed. In addition, a general method which will produce an arbitrary number of periodic orbits of shapes of arbitrary complexity is also discussed. A more general second method is given to construct a differential equation which converges to a fixed given finite set of equilibria. This technique is much more general in that it allows this set of equilibria to have any of a large class of indices which are consistent with the Morse Inequalities. It is clear that this class is not universal, because there is a large class of additional vector fields with convergent dynamics which cannot be constructed by the above method. The easiest way to see this is to enumerate the set of Morse indices which can be obtained by the above method and compare this class with the class of Morse indices of arbitrary differential equations with convergent dynamics. The former set of indices are a proper subclass of the latter, therefore, the above construction cannot be universal. In general, it is a difficult open problem to construct a specific example of a differential equation with a given fixed set of equilibria, permissible Morse indices, and permissible connections between stable and unstable manifolds. A strict Lyapunov function is given for this second case as well. This strict Lyapunov function as above enables construction of a large class of examples consistent with these more complicated dynamics and indices. The determination of all the basins of attraction in the general case for these systems is also difficult and open.

How Spinal Neural Networks Reduce Discrepancies between Motor Intention and Motor Realization

This paper attempts a rational, step-by-step reconstruction of many aspects of the mammalian neural circuitry known to be involved in the spinal cord's regulation of opposing muscles acting on skeletal segments. Mathematical analyses and local circuit simulations based on neural membrane equations are used to clarify the behavioral function of five fundamental cell types, their complex connectivities, and their physiological actions. These cell types are: α-MNs, γ-MNs, IaINs, IbINs, and Renshaw cells. It is shown that many of the complexities of spinal circuitry are necessary to ensure near invariant realization of motor intentions when descending signals of two basic types independently vary over large ranges of magnitude and rate of change. Because these two types of signal afford independent control, or Factorization, of muscle LEngth and muscle TEnsion, our construction was named the FLETE model (Bullock and Grossberg, 1988b, 1989). The present paper significantly extends the range of experimental data encompassed by this evolving model.

The Synthesis of Arbitrary Stable Dynamics in Non-linear Neural Networks II: Feedback and Universality

We wish to construct a realization theory of stable neural networks and use this theory to model the variety of stable dynamics apparent in natural data. Such a theory should have numerous applications to constructing specific artificial neural networks with desired dynamical behavior. The networks used in this theory should have well understood dynamics yet be as diverse as possible to capture natural diversity. In this article, I describe a parameterized family of higher order, gradient-like neural networks which have known arbitrary equilibria with unstable manifolds of known specified dimension. Moreover, any system with hyperbolic dynamics is conjugate to one of these systems in a neighborhood of the equilibrium points. Prior work on how to synthesize attractors using dynamical systems theory, optimization, or direct parametric. fits to known stable systems, is either non-constructive, lacks generality, or has unspecified attracting equilibria. More specifically, We construct a parameterized family of gradient-like neural networks with a simple feedback rule which will generate equilibrium points with a set of unstable manifolds of specified dimension. Strict Lyapunov functions and nested periodic orbits are obtained for these systems and used as a method of synthesis to generate a large family of systems with the same local dynamics. This work is applied to show how one can interpolate finite sets of data, on nested periodic orbits.

Genetically Generated Neural Networks I: Representational Effects

This paper studies several applications of genetic algorithms (GAs) within the neural networks field. After generating a robust GA engine, the system was used to generate neural network circuit architectures. This was accomplished by using the GA to determine the weights in a fully interconnected network. The importance of the internal genetic representation was shown by testing different approaches. The effects in speed of optimization of varying the constraints imposed upon the desired network were also studied. It was observed that relatively loose constraints provided results comparable to a fully constrained system. The type of neural network circuits generated were recurrent competitive fields as described by Grossberg (1982).

Genetically Generated Neural Networks II: Searching for an Optimal Representation

Genetic Algorithms (GAs) make use of an internal representation of a given system in order to perform optimization functions. The actual structural layout of this representation, called a genome, has a crucial impact on the outcome of the optimization process. The purpose of this paper is to study the effects of different internal representations in a GA, which generates neural networks. A second GA was used to optimize the genome structure. This structure produces an optimized system within a shorter time interval.

Eliminating Redundant Training Data Using Unsupervised Clustering Techniques

Training data for supervised learning neural networks can be clustered such that the input/output pairs in each cluster are redundant. Redundant training data can adversely affect training time. In this paper we apply two clustering algorithms, ART2 -A and the Generalized Equality Classifier, to identify training data clusters and thus reduce the training data and training time. The approach is demonstrated for a high dimensional nonlinear continuous time mapping. The demonstration shows six-fold decrease in training time at little or no loss of accuracy in the handling of evaluation data.

ART-EMAP: A Neural Network Architecture for Object Recognition by Evidence Accumulation

A new neural network architecture is introduced for the recognition of pattern classes after supervised and unsupervised learning. Applications include spatio-temporal image understanding and prediction and 3-D object recognition from a series of ambiguous 2-D views. The architecture, called ART-EMAP, achieves a synthesis of adaptive resonance theory (ART) and spatial and temporal evidence integration for dynamic predictive mapping (EMAP). ART-EMAP extends the capabilities of fuzzy ARTMAP in four incremental stages. Stage 1 introduces distributed pattern representation at a view category field. Stage 2 adds a decision criterion to the mapping between view and object categories, delaying identification of ambiguous objects when faced with a low confidence prediction. Stage 3 augments the system with a field where evidence accumulates in medium-term memory (MTM). Stage 4 adds an unsupervised learning process to fine-tune performance after the limited initial period of supervised network training. Each ART-EMAP stage is illustrated with a benchmark simulation example, using both noisy and noise-free data. A concluding set of simulations demonstrate ART-EMAP performance on a difficult 3-D object recognition problem.

Fuzzy ARTMAP: A Neural Network Architecture for Incremental Supervised Learning of Analog Multidimensional Maps

A new neural network architecture is introduced for incremental supervised learning of recognition categories and multidimensional maps in response to arbitrary sequences of analog or binary input vectors. The architecture, called Fuzzy ARTMAP, achieves a synthesis of fuzzy logic and Adaptive Resonance Theory (ART) neural networks by exploiting a close formal similarity between the computations of fuzzy subsethood and ART category choice, resonance, and learning. Fuzzy ARTMAP also realizes a new Minimax Learning Rule that conjointly minimizes predictive error and maximizes code compression, or generalization. This is achieved by a match tracking process that increases the ART vigilance parameter by the minimum amount needed to correct a predictive error. As a result, the system automatically learns a minimal number of recognition categories, or "hidden units", to met accuracy criteria. Category proliferation is prevented by normalizing input vectors at a preprocessing stage. A normalization procedure called complement coding leads to a symmetric theory in which the MIN operator (Λ) and the MAX operator (v) of fuzzy logic play complementary roles. Complement coding uses on-cells and off-cells to represent the input pattern, and preserves individual feature amplitudes while normalizing the total on-cell/off-cell vector. Learning is stable because all adaptive weights can only decrease in time. Decreasing weights correspond to increasing sizes of category "boxes". Smaller vigilance values lead to larger category boxes. Improved prediction is achieved by training the system several times using different orderings of the input set. This voting strategy can also be used to assign probability estimates to competing predictions given small, noisy, or incomplete training sets. Four classes of simulations illustrate Fuzzy ARTMAP performance as compared to benchmark back propagation and genetic algorithm systems. These simulations include (i) finding points inside vs. outside a circle; (ii) learning to tell two spirals apart; (iii) incremental approximation of a piecewise continuous function; and (iv) a letter recognition database. The Fuzzy ARTMAP system is also compared to Salzberg's NGE system and to Simpson's FMMC system.

AIRSTREAM: A Neural Network Model of Auditory Scene Analysis and Source Segregation

Multiple sound sources often contain harmonics that overlap and may be degraded by environmental noise. The auditory system is capable of teasing apart these sources into distinct mental objects, or streams. Such an "auditory scene analysis" enables the brain to solve the cocktail party problem. A neural network model of auditory scene analysis, called the AIRSTREAM model, is presented to propose how the brain accomplishes this feat. The model clarifies how the frequency components that correspond to a give acoustic source may be coherently grouped together into distinct streams based on pitch and spatial cues. The model also clarifies how multiple streams may be distinguishes and seperated by the brain. Streams are formed as spectral-pitch resonances that emerge through feedback interactions between frequency-specific spectral representaion of a sound source and its pitch. First, the model transforms a sound into a spatial pattern of frequency-specific activation across a spectral stream layer. The sound has multiple parallel representations at this layer. A sound's spectral representation activates a bottom-up filter that is sensitive to harmonics of the sound's pitch. The filter activates a pitch category which, in turn, activate a top-down expectation that allows one voice or instrument to be tracked through a noisy multiple source environment. Spectral components are suppressed if they do not match harmonics of the top-down expectation that is read-out by the selected pitch, thereby allowing another stream to capture these components, as in the "old-plus-new-heuristic" of Bregman. Multiple simultaneously occuring spectral-pitch resonances can hereby emerge. These resonance and matching mechanisms are specialized versions of Adaptive Resonance Theory, or ART, which clarifies how pitch representations can self-organize durin learning of harmonic bottom-up filters and top-down expectations. The model also clarifies how spatial location cues can help to disambiguate two sources with similar spectral cures. Data are simulated from psychophysical grouping experiments, such as how a tone sweeping upwards in frequency creates a bounce percept by grouping with a downward sweeping tone due to proximity in frequency, even if noise replaces the tones at their interection point. Illusory auditory percepts are also simulated, such as the auditory continuity illusion of a tone continuing through a noise burst even if the tone is not present during the noise, and the scale illusion of Deutsch whereby downward and upward scales presented alternately to the two ears are regrouped based on frequency proximity, leading to a bounce percept. Since related sorts of resonances have been used to quantitatively simulate psychophysical data about speech perception, the model strengthens the hypothesis the ART-like mechanisms are used at multiple levels of the auditory system. Proposals for developing the model to explain more complex streaming data are also provided.

Adaptive Neural Models of Queuing and Timing in Fluent Action

Temporal structure in skilled, fluent action exists at several nested levels. At the largest scale considered here, short sequences of actions that are planned collectively in prefrontal cortex appear to be queued for performance by a cyclic competitive process that operates in concert with a parallel analog representation that implicitly specifies the relative priority of elements of the sequence. At an intermediate scale, single acts, like reaching to grasp, depend on coordinated scaling of the rates at which many muscles shorten or lengthen in parallel. To ensure success of acts such as catching an approaching ball, such parallel rate scaling, which appears to be one function of the basal ganglia, must be coupled to perceptual variables, such as time-to-contact. At a fine scale, within each act, desired rate scaling can be realized only if precisely timed muscle activations first accelerate and then decelerate the limbs, to ensure that muscle length changes do not under- or over-shoot the amounts needed for the precise acts. Each context of action may require a much different timed muscle activation pattern than similar contexts. Because context differences that require different treatment cannot be known in advance, a formidable adaptive engine-the cerebellum-is needed to amplify differences within, and continuosly search, a vast parallel signal flow, in order to discover contextual "leading indicators" of when to generate distinctive parallel patterns of analog signals. From some parts of the cerebellum, such signals controls muscles. But a recent model shows how the lateral cerebellum, such signals control muscles. But a recent model shows how the lateral cerebellum may serve the competitive queuing system (in frontal cortex) as a repository of quickly accessed long-term sequence memories. Thus different parts of the cerebellum may use the same adaptive engine system design to serve the lowest and the highest of the three levels of temporal structure treated. If so, no one-to-one mapping exists between levels of temporal structure and major parts of the brain. Finally, recent data cast doubt on network-delay models of cerebellar adaptive timing.