S-TREE: self-organizing trees for data clustering and online vector quantization.

This paper introduces S-TREE (Self-Organizing Tree), a family of models that use unsupervised learning to construct hierarchical representations of data and online tree-structured vector quantizers. The S-TREE1 model, which features a new tree-building algorithm, can be implemented with various cost functions. An alternative implementation, S-TREE2, which uses a new double-path search procedure, is also developed. The performance of the S-TREE algorithms is illustrated with data clustering and vector quantization examples, including a Gauss-Markov source benchmark and an image compression application. S-TREE performance on these tasks is compared with the standard tree-structured vector quantizer (TSVQ) and the generalized Lloyd algorithm (GLA). The image reconstruction quality with S-TREE2 approaches that of GLA while taking less than 10% of computer time. S-TREE1 and S-TREE2 also compare favorably with the standard TSVQ in both the time needed to create the codebook and the quality of image reconstruction.

Neural-network models of learning and memory: leading questions and an emerging framework.

Real-time neural-network models provide a conceptual framework for formulating questions about the nature of cognition, an architectural framework for mapping cognitive functions to brain regions, a semantic framework for defining terms, and a computational framework for testing hypotheses. This article considers key questions about how a physical system might simultaneously support one-trial learning and lifetime memories, in the context of neural models that test possible solutions to the problems posed. Model properties point to partial answers, and model limitations lead to new questions. Placing individual system components in the context of a unified real-time network allows analysis to move from the level of neural processes, including learning laws and rules of synaptic transmission, to cognitive processes, including attention and consciousness.

Networks and groups within the genus Neisseria: analysis of argF, recA, rho, and 16S rRNA sequences from human Neisseria species.

To understand the pattern of nucleotide sequence variation among bacteria that frequently exchange chromosomal genes, we analyzed sequences of the recA, argF, and rho genes, as well as part of the small-subunit (16S) rRNA gene, from about 50 isolates of human commensal Neisseria species and the pathogenic N. meningitidis and N. gonorrhoeae. Almost all isolates of these species could be assigned to five phylogenetic groups that are found for all genes examined and generally are supported by high bootstrap values. In contrast, the phylogenetic relationships among groups varied according to the gene analyzed with notable incongruences involving N. cinerea and N. lactamica. Further analysis using split decomposition showed that for each gene, including 16S rRNA, the patterns of sequence divergence within N. meningitidis and closely related species were inconsistent with a bifurcating treelike phylogeny and better represented by an interconnected network. These data indicate that the human commensal Neisseria species can be separated into discrete groups of related species but that the relationships both within and among these groups, including those reconstructed using 16S rRNA, have been distorted by interspecies recombination events.

ARTMAP-IC and medical diagnosis: Instance counting and inconsistent cases.

For complex database prediction problems such as medical diagnosis, the ARTMAP-IC neural network adds distributed prediction and category instance counting to the basic fuzzy ARTMAP system. For the ARTMAP match tracking algorithm, which controls search following a predictive error, a new version facilitates prediction with sparse or inconsistent data. Compared to the original match tracking algorithm (MT+), the new algorithm (MT-) better approximates the real-time network differential equations and further compresses memory without loss of performance. Simulations examine predictive accuracy on four medical databases: Pima Indian diabetes, breast cancer, heart disease, and gall bladder removal. ARTMAP-IC results are equal to or better than those of logistic regression, K nearest neighbour (KNN), the ADAP preceptron, multisurface pattern separation, CLASSIT, instance-based (IBL), and C4. ARTMAP dynamics are fast, stable, and scalable. A voting strategy improves prediction by training the system several times on different orderings of an input set. Voting, instance counting, and distributed representations combine to form confidence estimates for competing predictions.

Ventricular outflow tract reconstructions with cryopreserved cardiac valve homografts. A single surgeon's 10-year experience.

OBJECTIVE: From January 1, 1985 through December 31, 1994, one surgeon implanted cryopreserved valved homografts into 149 patients--65 since December 1988. This latter series (II) was accomplished in a single hospital, facilitating patient follow-up with biannual echocardiograms. Analysis of these 65 patients is the primary focus of this report; the indications and early surgical results for the two parts of the series (I and II) are compared to assess the evolution of a single surgeon's use of homografts in a mixed pediatric and adult practice. METHODS: Fifty-one variables for each patient (series II) were entered into a computerized database and analyzed (multivariate and univariate) using SPSS 6.1 software (Statistical Products and Service Solutions, Chicago, IL). Cox proportional hazard model was used to identify the independent contribution of each variable for patient mortality and homograft failure. Cumulative survival estimates were made using Kaplan-Meier analysis. Homograft failure was defined as requirement for replacement or death. In series I, there were 41 left ventricular outflow tract (LVOT) reconstructions (31 adult) and 43 right ventricular outflow tract (RVOT) reconstructions (42 pediatric). In series II, there were 55 RVOT reconstructions (52 pediatric) and 10 LVOT reconstructions (7 adult). RESULTS: There were no technical surgical failures. Total surgical mortality rate was 6% (5/84) in series I (3 LVOT, 2 RVOT) and 15% (10/65) in series II (2 LVOT, 8 RVOT) (I vs. II NS; p = 0.11, two-tailed Fisher exact test). By the Cox analysis, only age < 2 years (p < 0.03) and cross-clamp time > 120 minutes (p < 0.05) were significant predictors for death. Age-based survival curves were compared in a sequential bivariate analyses (log rank test) and age < 2 years again was a significant predictor of decreased patient survival (p < 0.006). Actuarial freedom from patient death or reoperation for homograft failure was 82% +/- 7% at 1000 days and 77% +/- 10% at 2000 days. Three patients required re-replacement for homograft failure (5.4%); one of these patients died. The only significant predictor of homograft failure was postoperative endocarditis (p < 0.05). Homograft performance was evaluated by an extensive echocardiography protocol: in surviving patients and homografts, three valved conduits were judged to have severely impaired performance (stenosis or regurgitation), awaiting surgical replacement for a putative total homograft-related structural failures rate of 11% at 5 1/2 years. CONCLUSIONS: Comparisons of series I and II shows, in one surgeon's practice, an evolution away from use of cryopreserved homografts for LVOT reconstructions except when needed for destructive bacterial endocarditis or complex congenital anatomy. Homograft efficacy and durability were similar in RVOT and LVOT positions, with 78.5% of patients surviving at 5 1/2 years; in surviving patients, 89% of homografts have continued to function well. Homografts are not immune to prosthetic bacterial endocarditis, and its occurrence is associated with accelerated deterioration. Cryopreserved homograft valves are an imperfect but satisfactory biological material for specific ventricular outflow reconstructions.

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 3D object recognition from a series of ambiguous 2D 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. 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 fuzzy ARTMAP nonparametric probability estimator for nonstationary pattern recognition problems.

An incremental, nonparametric probability estimation procedure using the fuzzy ARTMAP (adaptive resonance theory-supervised predictive mapping) neural network is introduced. In the slow-learning mode, fuzzy ARTMAP searches for patterns of data on which to build ever more accurate estimates. In max-nodes mode, the network initially learns a fixed number of categories, and weights are then adjusted gradually.

Normal and amnesic learning, recognition and memory by a neural model of cortico-hippocampal interactions.

The processes by which humans and other primates learn to recognize objects have been the subject of many models. Processes such as learning, categorization, attention, memory search, expectation and novelty detection work together at different stages to realize object recognition. In this article, Gail Carpenter and Stephen Grossberg describe one such class of model (Adaptive Resonance Theory, ART) and discuss how its structure and function might relate to known neurological learning and memory processes, such as how infero-temporal cortex can recognize both specialized and abstract information, and how medial temporal amnesia might be caused by lesions in the hippocampal formation. This model also suggests how hippocampal and inferotemporal processing might be linked during recognition learning.

Fuzzy ARTMAP: A neural network architecture for incremental supervised learning of analog multidimensional maps.

A 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, which may represent fuzzy or crisp sets of features. 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. Four classes of simulation illustrated fuzzy ARTMAP performance in relation to benchmark backpropagation and generic algorithm systems. These simulations include finding points inside versus outside a circle, learning to tell two spirals apart, incremental approximation of a piecewise-continuous function, and a letter recognition database. The fuzzy ARTMAP system is also compared with Salzberg's NGE systems and with Simpson's FMMC system.

Air hygiene in a pullet house: effects of air filtration on aerial pollutants measured in vivo and in vitro.

1. The natural history of air hygiene in a pullet house was assessed at three-weekly intervals using a combination of in vitro and in vivo assays. The performance of an internal air filter was also examined as an experimental technique for providing clean air. 2. Overall, air hygiene was poor by comparison with occupational standards for human health. The mass concentrations of respirable and inspirable dust were 1.4 and 11.3 mg/m3 compared to human exposure limits of 5 and 10 mg/m3 respectively. The concentration of ammonia was typically about 20 ppm. 3. The majority (greater than 99%) of airborne particles were non-viable. Commensal bacteria from the skin were the most numerous airborne bacteria. Scopulariopsis and Aspergillus spp. were the most prevalent fungi recovered from the air and bird's lungs respectively. The concentrations of airborne and lung fungi were positively correlated with ammonia concentrations. 4. The differences between in vivo and in vitro assays of airborne microorganisms suggest that an aerosol sampler should be devised which better mimics the physical and biochemical environment of the respiratory tract.

Neural networks: introduction to the 1 December 1987 issue of Applied Optics.

In recent decades, research on the structure and function of mind and brain has led to the development of a powerful new paradigm which is now being applied and adapted to a variety of problems in many disciplines. This paradigm is based on the mathematical theory of neural networks. Neural network theory includes formal architectures comprised of a large number of highly interconnected neuronlike processing elements having adjustable or adaptive interconnection strengths. Some twenty-one papers in the 1 December 1987 issue of Applied Optics bring together the work of scientists who are developing the basic theory as well as its technological uses.

ART 2: self-organization of stable category recognition codes for analog input patterns.

Adaptive resonance architectures are neural networks that self-organize stable pattern recognition codes in real-time in response to arbitrary sequences of input patterns. This article introduces ART 2, a class of adaptive resonance architectures which rapidly self-organize pattern recognition categories in response to arbitrary sequences of either analog or binary input patterns. In order to cope with arbitrary sequences of analog input patterns-ART 2 architectures embody solutions to a number of design principles, such as the stability-plasticity tradeoff, the search-direct access tradeoff, and the match-reset tradeoff. In these architectures, top-down learned expectation and matching mechanisms are critical in self-stabilizing the code learning process. A parallel search scheme updates itself adaptively as the learning process unfolds, and realizes a form of real-time hypothesis discovery, testing, learning, and recognition. After learning selfstabilizes, the search process is automatically disengaged. Thereafter input patterns directly access their recognition codes without any search. Thus recognition time for familiar inputs does not increase with the complexity of the learned code. A novel input pattern can directly access a category if it shares invariant properties with the set of familiar exemplars of that category. A parameter called the attentional vigilance parameter determines how fine the categories will be. If vigilance increases (decreases) due to environmental feedback, then the system automatically searches for and learns finer (coarser) recognition categories. Gain control parameters enable the architecture to suppress noise up to a prescribed level. The architecture's global design enables it to learn effectively despite the high degree of nonlinearity of such mechanisms.

A neural theory of circadian rhythms: split rhythms, after-effects and motivational interactions.

A neural theory of the circadian pacemaker within the hypothalamic suprachiasmatic nuclei (SCN) is used to explain parametric data about mammalian operant behavior. The intensity, duration, and patterning of ultradian activity-rest cycles and the duration of circadian periods due to parametric (LL) and nonparametric (LD) lighting regimes are simulated. Paradoxical data about split rhythms and after-effects are explained using homeostatic and nonhomeostatic neural mechanisms that modulate pacemaker activity. These modulatory mechanisms enable the pacemaker to adjust to pervasive changes in its lighting regime, as during the passage of seasons, and to ultradian changes in internal metabolic conditions. The model circadian mechanisms are homologous to mechanisms that model hypothalamically mediated appetitive behaviors, such as eating. The theory thus suggests that both circadian and appetitive hypothalamic circuits are constructed from similar neural components. Mechanisms of transmitter habituation, opponent feedback interactions between on-cells and off-cells, homeostatic negative feedback, and conditioning are used in both the circadian and the appetitive circuits. Output from the SCN circadian pacemaker is assumed to modulate the sensitivity of the appetitive circuits to external and internal signals by controlling their level of arousal. Both underarousal and overarousal can cause abnormal behavioral syndromes whose properties have been found in clinical data. A model pacemaker can also be realized as an intracellular system.

A neural theory of circadian rhythms: Aschoff's rule in diurnal and nocturnal mammals.

A neural model of the suprachiasmatic nuclei suggests how behavioral activity, rest, and circadian period depend on light intensity in diurnal and nocturnal mammals. These properties are traced to the action of light input (external zeitgeber) and an activity-mediated fatigue signal (internal zeitgeber) on the circadian pacemaker. Light enhances activity of the diurnal model and suppresses activity of the nocturnal model. Fatigue suppresses activity in both diurnal and nocturnal models. The asymmetrical action of light and fatigue in diurnal vs. nocturnal models explains the more consistent adherence of nocturnal mammals to Aschoff's rule, the consistent adherence of both diurnal and nocturnal mammals to the circadian rule, and the tendency of nocturnal mammals to lose circadian rhythmicity at lower light levels than diurnal mammals. The fatigue signal is related to the sleep process S of Borbély (Hum. Neurobiol. 1: 195-204, 1982.) and contributes to the stability of circadian period. Two predictions follow: diurnal mammals obey Aschoff's rule less consistently during a self-selected light-dark cycle than in constant light, and if light level is increased enough during sleep in diurnal mammals to compensate for eye closure, then Aschoff's rule will hold more consistently. The results are compared with those of Enright's model.

Which behavior does the lamprey central motor program mediate?

The isolated lamprey spinal cord, when bathed in 2 millimolar D-glutamic acid, will generate a pattern of motor neuron discharge that has generally been assumed to represent the central motor program for swimming. Motion pictures of behaving lampreys were analyzed by a computer algorithm to estimate undulatory movement parameters that could be directly compared with those generated during D-glutamate--induced undulations. The D-glutamate--induced movement parameters were significantly different from those observed during normal behaviors, including swimming, but accurately predicted the undulations produced by spinally transected adult lampreys.

An analysis of the mammalian ventricular action potential.

A Hodgkin-Huxley model for ventricular excitation is abstracted from electrophysiological data. A singular perturbation analysis of the 8-dimensional phase portrait of the model characterizes the role of calcium during the plateau phase of the ventricular action potential and demonstrates how the calcium refractory period prevents tetanization.