J. David Schaffer

Evolving Spike Neural Network Sensors to Characterize the Alcoholic Brain Using Visually Evoked Response Potential

The electrical activity of the brain in response to a visual stimulus can be recorded using EEG. These signals are complex spatially-distributed time series. Here we investigate if it is possible to find hidden temporal patterns in these evoked electrical signals that could characterize the alcoholic brain. We have developed a technology for evolving spike neural network (SNN) sensors for detecting such hidden patterns in time-varying signals. The evolutionary computation involves a novel chromosome structure and a hybrid crossover operator for it. We have also developed a design rule for SNN-based temporal pattern detectors (TPD) that can detect a predefined inter-spike interval pattern in an incoming spike train. The design rule eliminates the need to tune the network parameters leaving only the design specifications to be learned. The primary goal of the evolutionary process is to select a set of EEG leads along with weights and to evolve the design specifications for the TPDs. After converting the composite EEG signal to a spike train, the TPDs are evaluated based on their ability to distinguish the alcoholic and the control cases. The early results suggest that this approach may be reliably used for characterizing the alcoholic brain.

Evolving spiking neural networks: A novel growth algorithm corrects the teacher

Spiking neural networks (SNNs) have generated considerable excitement because of their computational properties, believed to be superior to conventional von Neumann machines, and sharing properties with living brains. Yet progress building these systems has been limited because we lack a design methodology. We present a gene-driven network growth algorithm that enables a genetic algorithm (evolutionary computation) to generate and test SNNs. The genome length for this algorithm grows O(n) where n is the number of neurons; n is also evolved. The genome not only specifies the network topology, but all its parameters as well. In experiments, the algorithm discovered SNNs that effectively produce a robust spike bursting behavior given tonic inputs, an application suitable for central pattern generators. Even though evolution did not include perturbations of the input spike trains, the evolved networks showed remarkable robustness to such perturbations. On a second task, a sequence detector, several related discriminating designs were found, all made “errors” in that they fired when input spikes were simultaneous (i.e. not strictly in sequence), but not when they were out of sequence. They also fired when the sequence was too close for the teacher to have declared they were in sequence. That is, evolution produced these behaviors even though it was not explicitly rewarded for doing so. We are optimistic that this technology might be scaled up to produce robust SNN designs that humans would be hard pressed to produce.

GRNN Ensemble Classifier for Lung Cancer Prognosis Using Only Demographic and TNM features

Abstract Predicting the recurrence of non-small cell lung cancer remains a clinical challenge. The current best practice employs heuristic decisions based on the TNM classification scheme that many believe can be improved upon. Much research has recently been devoted to searching for gene signatures derived from gene expression microarrays for this challenge, but a consensus signature is still elusive. We present an approach to first create a benchmark for recurrence prediction based only upon gender, age and TNM features that uses several learning classifier induction methods and combines them into an ensemble using a recent extension of the general regression neural network. Using this approach on a pooled sample of 422 patients from two previously published studies (Shedden and Raponi), we demonstrate error rates in the low 20% for both false positives and negatives. Future work will focus on discovering if gene signatures can be discovered that can improve this performance.

Evolutionary computation with noise perturbation and cluster analysis to discover biomarker sets

Abstract In biomedical science, data mining techniques have been applied to extract statistically significant and clinically useful information from a given dataset. Finding biomarker gene sets for diseases can aid in understanding disease diagnosis, prognosis and therapy response. Gene expression microarrays have played an important role in such studies and yet, there have also been criticisms in their analysis. Analysis of these datasets presents the high risk of over-fitting (discovering spurious patterns) because of their feature-rich but case-poor nature. This paper describes a GA-SVM hybrid along with Gaussian noise perturbation (with a manual noise gain) to combat over-fitting; determine the strongest signal in the dataset; and discover stable biomarker sets. A colon cancer gene expression microarray dataset is used to show that the strongest signal in the data (optimal noise gain where a modest number of similar candidates emerge) can be found by a binary search. The diversity of candidates (measured by cluster analysis) is reduced by the noise perturbation, indicating some of the patterns are being eliminated (we hope mostly spurious ones). Initial biological validated has been tested and genes have different levels of significance to the candidates; although the discovered biomarker sets should be studied further to ascertain their biological significance and clinical utility. Furthermore, statistical validity displays that the strongest signal in the data is spurious and the discovered biomarker sets should be rejected.

A novel approach to signal classification with an application to identifying the alcoholic brain

Graphical abstractDisplay Omitted HighlightsEvolving temporal pattern detectors for signal classification.We make no assumptions regarding the spectral characteristics of the data.The classification accuracies are comparable to the conventional techniques.Located EEG-sensors that showed abnormal electrical behavior for the alcoholics.Evolutionary learning paradigm unified feature extraction and classification steps. We introduce a novel approach to signal classification based on evolving temporal pattern detectors (TPDs) that can find the occurrences of embedded temporal structures in discrete time signals and illustrate its application to characterizing the alcoholic brain using visually evoked response potentials. In contrast to conventional techniques used for most signal classification tasks, this approach unifies the feature extraction and classification steps. It makes no prior assumptions regarding the spectral characteristics of the data; it merely assumes that some temporal patterns exist that distinguish two classes of signals and therefore could be applied to new signal classification tasks where a body of prior work identifying important features does not exist. Evolutionary computation (EC) discovers a classifier by simply learning from the time series samples.The alcoholic classification (AC) problem consists of 2 sub-tasks, one spatial and one temporal: choosing a subset of electroencephalogram leads used to create a composite signal (the spatial task), and detecting temporal patterns in this signal that are more prevalent in the alcoholics than the controls (the temporal task). To accomplish this, a novel representation and crossover operator were devised that enable multiple feature subset tasks to be solved concurrently. Three TPD techniques are presented that differ in the mechanism by which partial credit is assigned to temporal patterns that deviate from the specified pattern. An EC approach is used for evolving a subset of sensors and the TPD specifications. We found evidence that partial credit does help evolutionary discovery. Regions on the skull of an alcoholic subject that produced abnormal electrical activity compared to the controls were located. These regions were consistent with prior findings in the literature. The classification accuracy was measured as the area under the receiver operator characteristic curve (ROC); the ROC area for the training set varied from 90.32% to 98.83% and for the testing set it varied from 87.17% to 95.9%.

Predicting with Confidence: Extensions to the GRNN Oracle Enabling Quantification of Confidence in Predictions☆

Abstract The GRNN oracle is an optimal estimator that provides the maximum likelihood unbiased estimate by combining a series of intelligent processing results, where those estimates with the smallest variance are weighted most highly. It is known that if the individual predictors in the ensemble are too similar, the oracle cannot provide much improvement. We have newly observed that if the predictions are characterized by class inhomogeneities, then the oracle can be limited in its ability to compensate. For some training cases, all models might provide incorrect predictions; let us call these cases “trouble makers.” To address this problem, the oracle theory was mathematically extended to provide estimates of the sensitivity of its predictions. These sensitivities provide a basis for declaring that certain of its predictions should be treated as untrustworthy. It then has information to flag them. This paper addresses that theoretical development and applies these extensions, to toy problems, with the future objective of application to real problems of detecting dementia / Alzheimers in speech patterns.

Perturbation and candidate analysis to combat overfitting of gene expression microarray data

Analysis of gene expression microarray datasets presents the high risk of over-fitting (spurious patterns) because of their feature-rich but case-poor nature. This paper describes our ongoing efforts to develop a method to combat over-fitting and determine the strongest signal in the dataset. A GA-SVM hybrid along with Gaussian noise (manual noise gain) is used to discover feature sets of minimal size that accurately classifies the cases under cross-validation. Initial results on a colorectal cancer dataset shows that the strongest signal (modest number of candidates) can be found by a binary search.

Evolving spiking neural networks: a novel growth algorithm exhibits unintelligent design

Spiking neural networks (SNNs) have drawn considerable excitement because of their computational properties, believed to be superior to conventional von Neumann machines, and sharing properties with living brains. Yet progress building these systems has been limited because we lack a design methodology. We present a gene-driven network growth algorithm that enables a genetic algorithm (evolutionary computation) to generate and test SNNs. The genome for this algorithm grows O(n) where n is the number of neurons; n is also evolved. The genome not only specifies the network topology, but all its parameters as well. Experiments show the algorithm producing SNNs that effectively produce a robust spike bursting behavior given tonic inputs, an application suitable for central pattern generators. Even though evolution did not include perturbations of the input spike trains, the evolved networks showed remarkable robustness to such perturbations. In addition, the output spike patterns retain evidence of the specific perturbation of the inputs, a feature that could be exploited by network additions that could use this information for refined decision making if required. On a second task, a sequence detector, a discriminating design was found that might be considered an example of “unintelligent design”; extra non-functional neurons were included that, while inefficient, did not hamper its proper functioning.

Investigating the GRNN Oracle as a Method for Combining Multiple Predictive Models of Colon Cancer Recurrence from Gene Microarrays

Abstract In previous work, we applied an advanced genetic algorithm method for feature subset selection combined with noise perturbation in an attempt to overcome the over-fitting that is typical with microarray datasets. The method was applied to a dataset from Moffitt Cancer Center and the clinical outcome to be predicted was cancer recurrence in less than 5 years. By its nature, the method yields multiple gene signatures, each as small as possible and often these signatures will share one or more genes. The question is how to combine the predictions from multiple predictors. In the previous work, we produced an ensemble prediction by a simple majority vote rule, and observed that performance on a validation set was considerably worse than on the learning set. Our conclusion was that the training and validation sets were not equally representative of the same population, and therefore could not provide reliable gene signatures. Here we report on an effort to apply a more sophisticated ensemble method, the Generalized Regression Neural network (GRNN) Oracle, but this did not allow us to reverse our original conclusion.

Modeling Low Reynolds Number Flows Driven by Forward-Propagating and Reflected Boundary Waves in Concentric Micro-Cylinders

We report on a computational model used to study the reversal of flow direction inside the annular region between concentric micro-cylinders filled with an incompressible Newtonian fluid. The flow is induced by boundary deformations on the inner and outer cylinder surfaces due to forward-propagating transverse waves and their reflections. This microfluidic transport mechanism is postulated as a vital pathway for removal of beta-amyloid from the brain along sub-millimeter cerebral arteries, and failure of this clearance is associated with Alzheimer’s disease. We show that the direction of this annular flow depends on superposition of the peristaltic waves and their reflection waves. A control volume analysis is developed to predict the transport characteristics and compared with numerical solutions of the Navier-Stokes equations. The identified set of microfluidic parameters that leads to a net reverse flow will aid biologists in understanding why an aging brain becomes prone to beta-amyloid accumulation.Copyright