A study on rule extraction from several combined neural networks

The problem of rule extraction from neural networks is NP-hard. This work presents a new technique to extract "if-then-else" rules from ensembles of DIMLP neural networks. Rules are extracted in polynomial time with respect to the dimensionality of the problem, the number of examples, and the size of the resulting network. Further, the degree of matching between extracted rules and neural network responses is 100%. Ensembles of DIMLP networks were trained on four data sets in the public domain. Extracted rules were on average significantly more accurate than those extracted from C4.5 decision trees.     

Knowledge extraction from evolving spiking neural networks with rank order population coding

This paper demonstrates how knowledge can be extracted from evolving spiking neural networks with rank order population coding. Knowledge discovery is a very important feature of intelligent systems. Yet, a disproportionally small amount of research is centered on the issue of knowledge extraction from spiking neural networks which are considered to be the third generation of artificial neural networks. The lack of knowledge representation compatibility is becoming a major detriment to end users of these networks. We show that a high-level knowledge can be obtained from evolving spiking neural networks. More specifically, we propose a method for fuzzy rule extraction from an evolving spiking network with rank order population coding. The proposed method was used for knowledge discovery on two benchmark taste recognition problems where the knowledge learnt by an evolving spiking neural network was extracted in the form of zero-order Takagi-Sugeno fuzzy IF-THEN rules.     

High-throughput ligand screening via preclustering and evolved neural networks

The pathway for novel lead drug discovery has many major deficiencies, the most significant of which is the immense size of small molecule diversity space. Methods that increase the search efficiency and/or reduce the size of the search space, increase the rate at which useful lead compounds are identified. Artificial neural networks optimized via evolutionary computation provide a cost and time-effective solution to this problem. Here, we present results that suggest preclustering of small molecules prior to neural network optimization is useful for generating models of quantitative structure-activity relationships for a set of HIV inhibitors. Using these methods, it is possible to prescreen compounds to separate active from inactive compounds or even actives and mildly active compounds from inactive compounds with high predictive accuracy while simultaneously reducing the feature space. It is also possible to identify "human interpretable" features from the best models that can be used for proposal and synthesis of new compounds in order to optimize potency and specificity.     

Rules extraction from neural networks applied to the prediction and recognition of prokaryotic promoters

Promoters are DNA sequences located upstream of the gene region and play a central role in gene expression. Computational techniques show good accuracy in gene prediction but are less successful in predicting promoters, primarily because of the high number of false positives that reflect characteristics of the promoter sequences. Many machine learning methods have been used to address this issue. Neural Networks (NN) have been successfully used in this field because of their ability to recognize imprecise and incomplete patterns characteristic of promoter sequences. In this paper, NN was used to predict and recognize promoter sequences in two data sets: (i) one based on nucleotide sequence information and (ii) another based on stability sequence information. The accuracy was approximately 80% for simulation (i) and 68% for simulation (ii). In the rules extracted, biological consensus motifs were important parts of the NN learning process in both simulations.     

A biologically motivated neural network for phase extraction from complex sounds

We demonstrate that natural acoustic signals like speech or music contain synchronous phase information across multiple frequency bands and show how to extract this information using a spiking neural network. This network model is motivated by common neurophysiological findings in the auditory brainstem and midbrain of several species. A computer simulation of the model was tested by applying spoken vowels and organ pipe tones. As expected, spikes occurred synchronously in the activated frequency bands. This phase information may be used for sound separation with one microphone or sound localization with two microphones.     

Recognizing recurrent neural networks (rRNN): Bayesian inference for recurrent neural networks

Recurrent neural networks (RNNs) are widely used in computational neuroscience and machine learning applications. In an RNN, each neuron computes its output as a nonlinear function of its integrated input. While the importance of RNNs, especially as models of brain processing, is undisputed, it is also widely acknowledged that the computations in standard RNN models may be an over-simplification of what real neuronal networks compute. Here, we suggest that the RNN approach may be made computationally more powerful by its fusion with Bayesian inference techniques for nonlinear dynamical systems. In this scheme, we use an RNN as a generative model of dynamic input caused by the environment, e.g. of speech or kinematics. Given this generative RNN model, we derive Bayesian update equations that can decode its output. Critically, these updates define a 'recognizing RNN' (rRNN), in which neurons compute and exchange prediction and prediction error messages. The rRNN has several desirable features that a conventional RNN does not have, e.g. fast decoding of dynamic stimuli and robustness to initial conditions and noise. Furthermore, it implements a predictive coding scheme for dynamic inputs. We suggest that the Bayesian inversion of RNNs may be useful both as a model of brain function and as a machine learning tool. We illustrate the use of the rRNN by an application to the online decoding (i.e. recognition) of human kinematics.     

Signal-to-signal neural networks for improved spike estimation from calcium imaging data

Spiking information of individual neurons is essential for functional and behavioral analysis in neuroscience research. Calcium imaging techniques are generally employed to obtain activities of neuronal populations. However, these techniques result in slowly-varying fluorescence signals with low temporal resolution. Estimating the temporal positions of the neuronal action potentials from these signals is a challenging problem. In the literature, several generative model-based and data-driven algorithms have been studied with varied levels of success. This article proposes a neural network-based signal-to-signal conversion approach, where it takes as input raw-fluorescence signal and learns to estimate the spike information in an end-to-end fashion. Theoretically, the proposed approach formulates the spike estimation as a single channel source separation problem with unknown mixing conditions. The source corresponding to the action potentials at a lower resolution is estimated at the output. Experimental studies on the spikefinder challenge dataset show that the proposed signal-to-signal conversion approach significantly outperforms state-of-the-art-methods in terms of Pearson’s correlation coefficient, Spearman’s rank correlation coefficient and yields comparable performance for the area under the receiver operating characteristics measure. We also show that the resulting system: (a) has low complexity with respect to existing supervised approaches and is reproducible; (b) is layer-wise interpretable, and (c) has the capability to generalize across different calcium indicators.     

Synchrony measures for biological neural networks

 Synchronous firing of a population of neurons has been observed in many experimental preparations; in addition, various mathematical neural network models have been shown, analytically or numerically, to contain stable synchronous solutions. In order to assess the level of synchrony of a particular network over some time interval, quantitative measures of synchrony are needed. We develop here various synchrony measures which utilize only the spike times of the neurons; these measures are applicable in both experimental situations and in computer models. Using a mathematical model of the CA3 region of the hippocampus, we evaluate these synchrony measures and compare them with pictorial representations of network activity. We illustrate how synchrony is lost and synchrony measures change as heterogeneity amongst cells increases. Theoretical expected values of the synchrony measures for different categories of network solutions are derived and compared with results of simulations. Received: 6 June 1994/Accepted in revised form: 13 January 1995     

A neural network for the Steiner minimal tree problem

The problem of finding the shortest tree connecting a set of points is called the Steiner minimal tree problem and is nearly three centuries old. It has applications in transportation, computer networks, agriculture, telephony, building layout and very large scale integrated circuit (VLSI) design, among others, and is known to be NP-complete. We propose a neural network which self-organizes to find a minimal tree. Solutions found by the network compare favourably with the best known or optimal results on test problems from the literature. To the best of our knowledge, the proposed network is the first neural-based solution to the problem. We show that the neural network has a built-in mechanism to escape local minima.     

Leveraging high-throughput screening data,deep neural networks,and conditional generative adversarial networks to advance predictive toxicology

There are currently 85,000 chemicals registered with the Environmental Protection Agency (EPA) under the Toxic Substances Control Act, but only a small fraction have measured toxicological data. To address this gap, high-throughput screening (HTS) and computational methods are vital. As part of one such HTS effort, embryonic zebrafish were used to examine a suite of morphological and mortality endpoints at six concentrations from over 1,000 unique chemicals found in the ToxCast library (phase 1 and 2). We hypothesized that by using a conditional generative adversarial network (cGAN) or deep neural networks (DNN), and leveraging this large set of toxicity data we could efficiently predict toxic outcomes of untested chemicals. Utilizing a novel method in this space, we converted the 3D structural information into a weighted set of points while retaining all information about the structure. In vivo toxicity and chemical data were used to train two neural network generators. The first was a DNN (Go-ZT) while the second utilized cGAN architecture (GAN-ZT) to train generators to produce toxicity data. Our results showed that Go-ZT significantly outperformed the cGAN, support vector machine, random forest and multilayer perceptron models in cross-validation, and when tested against an external test dataset. By combining both Go-ZT and GAN-ZT, our consensus model improved the SE, SP, PPV, and Kappa, to 71.4%, 95.9%, 71.4% and 0.673, respectively, resulting in an area under the receiver operating characteristic (AUROC) of 0.837. Considering their potential use as prescreening tools, these models could provide in vivo toxicity predictions and insight into the hundreds of thousands of untested chemicals to prioritize compounds for HT testing.     

Improved neuronal models for studying neural networks

Previous neuronal models used for the study of neural networks are considered. Equations are developed for a model which includes: 1) a normalized range of firing rates with decreased sensitivity at large excitatory or large inhibitory input levels, 2) a single rate constant for the increase in firing rate following step changes in the input, 3) one or more rate constants, as required to fit experimental data for the adaptation of firing rates to maintained inputs. Computed responses compare well with the types of neuronal responses observed experimentally. Depending on the parameters, overdamped increases and decreases, damped oscillatory or maintained oscillatory changes in firing rate are observed to step changes in the input. The integrodifferential equations describing the neuronal models can be represented by a set of first-order differential equations. Steady-state solutions for these equations can be obtained for constant inputs, as well as the stability of the solutions to small perturbations. The linear frequency response function is derived for sufficiently small time-varying inputs. The linear responses are also compared with the computed solutions for larger non-linear responses.     

Feed-forward neural networks for secondary structure prediction

A feed-forward neural network has been employed for protein secondary structure prediction. Attempts were made to improve on previous prediction accuracies using a hierarchical mixture of experts (HME). In this method input data are clustered and used to train a series of different networks. Application of an HME to the prediction of protein secondary structure is shown to provide no advantages over a single network. We have also tried various new input representations, chosen to incorporate the effect of residues a long distance away in the one-dimensional amino acid chain. Prediction accuracy using these methods is comparable to that achieved by other neural networks.^1–4     

Selecting neural networks for a committee decision

To improve recognition results, decisions of multiple neural networks can be aggregated into a committee decision. In contrast to the ordinary approach of utilizing all neural networks available to make a committee decision, we propose creating adaptive committees, which are specific for each input data point. A prediction network is used to identify classification neural networks to be fused for making a committee decision about a given input data point. The jth output value of the prediction network expresses the expectation level that the jth classification neural network will make a correct decision about the class label of a given input data point. The proposed technique is tested in three aggregation schemes, namely majority vote, averaging, and aggregation by the median rule and compared with the ordinary neural networks fusion approach. The effectiveness of the approach is demonstrated on two artificial and three real data sets.     

Population networks: a large-scale framework for modelling cortical neural networks

Artificial neural networks are usually built on rather few elements such as activation functions, learning rules, and the network topology. When modelling the more complex properties of realistic networks, however, a number of higher-level structural principles become important. In this paper we present a theoretical framework for modelling cortical networks at a high level of abstraction. Based on the notion of a population of neurons, this framework can accommodate the common features of cortical architecture, such as lamination, multiple areas and topographic maps, input segregation, and local variations of the frequency of different cell types (e.g., cytochrome oxidase blobs). The framework is meant primarily for the simulation of activation dynamics; it can also be used to model the neural environment of single cells in a multiscale approach. Received: 9 January 1996 / Accepted in revised form: 24 July 1996     

Sequential RAM-based neural networks: learnability, generalisation, knowledge extraction, and grammatical inference

A fundamental question in the field of artificial neural networks is what set of problems a given class of networks can perform (computability). Such a problem can be made less general, but no less important, by asking what these networks could learn by using a given training procedure (learnability). The basic purpose of this paper is to address the learnability problem. Specifically, it analyses the learnability of sequential RAM-based neural networks. The analytical tools used are those of Automata Theory. In this context, this paper establishes which class of problems and under what conditions such networks, together with their existing learning rules, can learn and generalize. This analysis also yields techniques for both extracting knowledge from and inserting knowledge into the networks. The results presented here, besides helping in a better understanding of the temporal behaviour of sequential RAM-based networks, could also provide useful insights for the integration of the symbolic/connectionist paradigms.     

Maternal serum alpha-fetoprotein screening for neural tube defects

Summary The basis of maternal serum alpha-fetoprotein (AFP)-screening for neural tube defects is discussed. A report is given of a large scale screening study in the Federal Republic of Germany combining the experiences in Giessen and Hannover on over 50,000 pregnant women, about evently distributed among both centers. Published and known forthcoming data from other low incidence populations, particularly of European countries, are reviewed briefly. The conclusion is reached that general screening could effectively be instituted and in the final result should also be cost-beneficial.     

The extraction of neural strategies from the surface EMG.

This brief review examines some of the methods used to infer central control strategies from surface electromyogram (EMG) recordings. Among the many uses of the surface EMG in studying the neural control of movement, the review critically evaluates only some of the applications. The focus is on the relations between global features of the surface EMG and the underlying physiological processes. Because direct measurements of motor unit activation are not available and many factors can influence the signal, these relations are frequently misinterpreted. These errors are compounded by the counterintuitive effects that some system parameters can have on the EMG signal. The phenomenon of crosstalk is used as an example of these problems. The review describes the limitations of techniques used to infer the level of muscle activation, the type of motor unit recruited, the upper limit of motor unit recruitment, the average discharge rate, and the degree of synchronization between motor units. Although the global surface EMG is a useful measure of muscle activation and assessment, there are limits to the information that can be extracted from this signal.