NCTA 2015 Abstracts


Full Papers
Paper Nr: 6
Title:

A Comparison of Learning Rules for Mixed Order Hyper Networks

Authors:

Kevin Swingler

Abstract: A mixed order hyper network (MOHN) is a neural network in which weights can connect any number of neurons, rather than the usual two. MOHNs can be used as content addressable memories with higher capacity than standard Hopfield networks. They can also be used for regression, clustering, classification, and as fitness models for use in heuristic optimisation. This paper presents a set of methods for estimating the values of the weights in a MOHN from training data. The different methods are compared to each other and to a standard MLP trained by back propagation and found to be faster to train than the MLP and more reliable as the error function does not contain local minima.

Paper Nr: 14
Title:

Neurosolver Learning to Solve Towers of Hanoi Puzzles

Authors:

Andrzej Bieszczad and Skyler Kuchar

Abstract: Neurosolver is a neuromorphic planner and a general problem solving (GPS) system. To acquire its problem solving capability, Neurosolver uses a structure similar to the columnar organization of the cortex of the brain and a notion of place cells. The fundamental idea behind Neurosolver is to model world using a state space paradigm, and then use the model to solve problems presented as a pair of two states of the world: the current state and the desired (i.e., goal) state. Alternatively, the current state may be known (e.g., through the use of sensors), so the problem is fully expressed by stating just the goal state. Mechanically, Neurosolver works as a memory recollection system in which training samples are given as sequences of states of the subject system. Neurosolver generates a collection of interconnected nodes (inspired by cortical columns), each of which represents a single point in the problem state space, with the connections representing state transitions. A connection map between states is generated during training, and using this learned memory information, Neurosolver is able to construct a path from its current state, to the goal state for each such pair for which a transitions is possible at all. In this paper we show that Neurosolver is capable of acquiring from scratch the complete knowledge necessary to solve any puzzle for a given Towers of Hanoi configuration.

Paper Nr: 15
Title:

Mining Significant Frequent Patterns in Parallel Episodes with a Graded Notion of Synchrony and Selective Participation

Authors:

Salatiel Ezennaya-Gomez and Christian Borgelt

Abstract: We consider the task of finding frequent parallel episodes in parallel point processes (or event sequences), allowing for imprecise synchrony of the events constituting occurrences (temporal imprecision) as well as incomplete occurrences (selective participation). The temporal imprecision problem is tackled by frequent pattern mining using a graded notion of synchrony that captures both the number of instances of a pattern as well as the precision of synchrony of its events. To cope with selective participation, a reduction sequence of items (or event types) is formed based on found frequent patterns and guided by pattern overlap. We evaluate the performance of this method on a large number of data sets with injected parallel episodes. We demonstrate that, in contrast to binary synchrony where it pays to consider the pattern instances, graded synchrony performs better with a pattern-based scheme than with an instance-based one, thus simplifying the procedure.

Paper Nr: 18
Title:

A Heteroassociative Learning Model Robust to Interference

Authors:

Randa Kassab and Frédéric Alexandre

Abstract: Neuronal models of associative memories are recurrent networks able to learn quickly patterns as stable states of the network. Their main acknowledged weakness is related to catastrophic interference when too many or too close examples are stored. Based on biological data we have recently proposed a model resistant to some kinds of interferences related to heteroassociative learning. In this paper we report numerical experiments that highlight this robustness and demonstrate very good performances of memorization. We also discuss convergence of interests for such an adaptive mechanism for biological modeling and information processing in the domain of machine learning.

Paper Nr: 19
Title:

Control of Three-Phase Grid-Connected Microgrids using Artificial Neural Networks

Authors:

Shuhui Li, Xingang Fu, Ishan Jaithwa, Eduardo Alonso, Michael Fairbank and Donald C. Wunsch

Abstract: A microgrid consists of a variety of inverter-interfaced distributed energy resources (DERs). A key issue is how to control DERs within the microgrid and how to connect them to or disconnect them from the microgrid quickly. This paper presents a strategy for controlling inverter-interfaced DERs within a microgrid using an artificial neural network, which implements a dynamic programming algorithm and is trained with a new Levenberg-Marquardt backpropagation algorithm. Compared to conventional control methods, our neural network controller exhibits fast response time, low overshoot, and, in general, the best performance. In particular, the neural network controller can quickly connect or disconnect inverter-interfaced DERs without the need for a synchronization controller, efficiently track fast-changing reference commands, tolerate system disturbances, and satisfy control requirements at grid-connected mode, islanding mode, and their transition.

Paper Nr: 22
Title:

An Extended Q Learning System with Emotion State to Make Up an Agent with Individuality

Authors:

Masanao Obayashi, Shunsuke Uto, Takashi Kuremoto, Shingo Mabu and Kunikazu Kobayashi

Abstract: Recently, researches for the intelligent robots incorporating knowledge of neuroscience have been actively carried out. In particular, a lot of researchers making use of reinforcement learning have been seen, especially, "Reinforcement learning methods with emotions", that has already proposed so far, is very attractive method because it made us possible to achieve the complicated object, which could not be achieved by the conventional reinforcement learning method, taking into account of emotions. In this paper, we propose an extended reinforcement (Q) learning system with amygdala (emotion) models to make up individual emotions for each agent. In addition, through computer simulations that the proposed method is applied to the goal search problem including a variety of distinctive solutions, it finds that each agent is able to have each individual solution.

Paper Nr: 30
Title:

Sources Separation of Mono Signal based on Convolutive NMF

Authors:

Giovanni Costantini, Massimiliano Todisco and Giovanni Saggio

Abstract: In many applications such as music transcription, audio forensics, speech denoising, it is needed to decompose a mono recording into its respective sources. These techniques are usually referred to as blind source separation (BSS). Recently, one of the techniques used in BSS is non-negative matrix factorization (NMF) both in supervised and unsupervised learning method. In this paper we focus on convolutive NMF algorithms to evaluate the performance of BSS in which supervised mode is used. The results on music mixtures of the MASS database based on signal to distortion ratio (SDR) and signal to artefact ratio (SAR) show that the proposed system perform a good reconstruction of sources signal.

Short Papers
Paper Nr: 4
Title:

Time Series Forecasting using Clustering with Periodic Pattern

Authors:

Jan Kostrzewa

Abstract: Time series forecasting have attracted a great deal of attention from various research communities. One of the method which improves accuracy of forecasting is time series clustering. The contribution of this work is a new method of clustering which relies on finding periodic pattern by splitting the time series into two subsequences (clusters) with lower potential error of prediction then whole series. Having such subsequences we predict their values separately with methods customized to the specificities of the subsequences and then merge results according to the pattern and obtain prediction of original time series. In order to check efficiency of our approach we perform analysis of various artificial data sets. We also present a real data set for which application of our approach gives more then 300% improvement in accuracy of prediction. We show that in artificially created series we obtain even more pronounced accuracy improvement. Additionally our approach can be use to noise filtering. In our work we consider noise of a periodic repetitive pattern and we present simulation where we find correct series from data where 50% of elements is random noise.

Paper Nr: 5
Title:

Predicting Flight Departure Delay at Porto Airport: A Preliminary Study

Authors:

Hugo Alonso and António Loureiro

Abstract: Managing an airport is very complex. Decisions are often based on common sense and influence several variables, such as flight delay. This paper considers the problem of predicting flight departure delay at Porto Airport. As far as we know, this the first study on the subject. The problem is treated as an ordinal classification task and a suitable approach, based on the so-called unimodal model, is used to predict the delay. The unimodal model is implemented using neural networks and, for comparison purposes, also using trees.

Paper Nr: 8
Title:

Pedestrian Action Prediction using Static Image Feature

Authors:

Kenji Nishida, Takumi Kobayashi, Taro Iwamoto and Shinya Yamasaki

Abstract: In this study, we propose a method to predict how the target object move (run or walk) in the future by using only appearance-based image features. Such kind of motion prediction significantly contributes to intelligent braking system in cars; by knowing that the objects will run in several seconds such as in crossing streets, the car can start to brake in advance, which effectively reduces the risk for crash accidents. In the proposed method, we empirically evaluate which frames preceding the target action, 'running' in this case, are effective for predicting it in the framework of feature selection. By using the most effective frames at which the image features are extracted, we can build the action prediction method. In the experiments, those frames are found around 0.37 second before running action and we also show that they are closely related to human motion phases from walking to running from the viewpoint of biomechanics.

Paper Nr: 11
Title:

Machine Learning Techniques and the Existence of Variant Processes in Humans Declarative Memory

Authors:

Alex Frid, Hananel Hazan, Ester Koilis, Larry M. Manevitz, Maayan Merhav and Gal Star

Abstract: This work uses supervised machine learning methods over fMRI brain scans to establish the existence of two different encoding procedures for human declarative memory. Declarative knowledge refers to the memory for facts and events and initially depends on the hippocampus. Recent studies which used patients with hippocampal lesions and neuroimaging data, suggested the existence of an alternative process to form declarative memories. This process is triggered by learning mechanism called "Fast Mapping (FM)", as opposed to the 'standard' "Explicit Encoding (EE)" learning procedure. The present work gives a clear biomarker on the existence of two distinct encoding procedures as we can accurately predict which of the processes is being used directly from voxel activity in fMRI scans. The scans are taken during retrieval of information wherein the tasks are identical regardless of which procedure was used for acquisition and by that reflect conclusive prediction. This is an identification of a more subtle cognitive task than direct perceptual cognitive tasks as it requires some encoding and processing in the brain.

Paper Nr: 13
Title:

Gaussian Nonlinear Line Attractor for Learning Multidimensional Data

Authors:

Theus H. Aspiras, Vijayan K. Asari and Wesam Sakla

Abstract: The human brain’s ability to extract information from multidimensional data modeled by the Nonlinear Line Attractor (NLA), where nodes are connected by polynomial weight sets. Neuron connections in this architecture assumes complete connectivity with all other neurons, thus creating a huge web of connections. We envision that each neuron should be connected to a group of surrounding neurons with weighted connection strengths that reduces with proximity to the neuron. To develop the weighted NLA architecture, we use a Gaussian weighting strategy to model the proximity, which will also reduce the computation times significantly. Once all data has been trained in the NLA network, the weight set can be reduced using a locality preserving nonlinear dimensionality reduction technique. By reducing the weight sets using this technique, we can reduce the amount of outputs for recognition tasks. An appropriate distance measure can then be used for comparing testing data and the trained data when processed through the NLA architecture. It is observed that the proposed GNLA algorithm reduces training time significantly and is able to provide even better recognition using fewer dimensions than the original NLA algorithm. We have tested this algorithm and showed that it works well in different datasets, including the EO Synthetic Vehicle database and the Sheffield face database.

Paper Nr: 16
Title:

Investment Support System using the EVOLINO Recurrent Neural Network Ensemble

Authors:

Algirdas Maknickas and Nijolė Maknickienė

Abstract: The chaotic and largely unpredictable conditions that prevail in exchange markets are of considerable interest to speculators because of the potential for profit. The creation and development of a support system using artificial intelligence algorithms provides new opportunities for investors in financial markets. Therefore, the authors have developed a support system that processes historical data, makes predictions using an ensemble of EVOLINO recurrent neural networks, assesses these predictions using a composition of high-low distributions, selects an orthogonal investment portfolio, and verifies the outcome on the real market. The support system requires multi-core hardware resources to allow for timely data processing using an MPI library-based parallel computation approach. A comparison of daily and weekly predictions reveals that weekly forecasts are less accurate than daily predictions, but are still accurate enough to trade successfully on the currency markets. Information obtained from the support system gives investors an advantage over uninformed market players in making investment decisions.

Paper Nr: 26
Title:

SO(2) Approximate Identity Neural Networks are Universal Approximators

Authors:

Saeed Panahian Fard and Zarita Zainuddin

Abstract: The idea of approximation functions on the rotation group has important applications in many fields of science and engineering. This study is devoted to explore the universal approximation capability of a class of three layer feedforward artificial neural networks on the special orthogonal rotation group SO(2). To do this end, we propose the concept of SO(2) approximate identity. Moreover, we prove a theorem that provides a connection between SO(2) approximate identity and uniform convergence in the space of continuous functions on the rotation group SO(2). Furthermore, we apply this theorem to set a main theorem. The main theorem shows that three layer feedforward SO(2) approximate identity neural networks are universal approximators in the space of continuous functions on the rotation group SO(2). The construction of the proof of the main theorem utilizes a method based on the notion of epsilon-net.

Paper Nr: 27
Title:

Diffusion Bases Dimensionality Reduction

Authors:

Alon Schclar and Amir Averbuch

Abstract: The overflow of data is a critical contemporary challenge in many areas such as hyper-spectral sensing, information retrieval, biotechnology, social media mining, classification etc. It is usually manifested by a high-dimensional representation of data observations. In most cases, the information that is inherent in highdimensional datasets is conveyed by a small number of parameters that correspond to the actual degrees of freedom of the dataset. In order to efficiently process the dataset, one needs to derive these parameters by embedding the dataset into a low-dimensional space. This process is commonly referred to as dimensionality reduction or feature extraction. We present a novel algorithm for dimensionality reduction – diffusion bases – which explores the connectivity among the coordinates of the data and is dual to the diffusion maps algorithm. The algorithm reduces the dimensionality of the data while maintaining the coherency of the information that is conveyed by the data.

Paper Nr: 29
Title:

Inedited SVM Application to Automatically Tracking and Recognizing Arm-and-Hand Visual Signals to Aircraft

Authors:

Giovanni Saggio, Francesco Cavrini and Franco Di Paolo

Abstract: An electronic demonstrator was designed and developed to automatically interpret the signalman’s arm-and-hand visual signals. It was based on an “extended” sensory glove, which is a glove equipped with sensors to measure fingers/wrist/forearm movements, an electronic circuitry to acquire/condition/feed measured data to a personal computer, SVM based routines to classify the visual signals, and a graphical interface to represent classified data. The aim was to furnish to the Italian Aircraft Force a tool for ground-to-ground or ground-to-air communication, which can be independent from the full view of the vehicle drivers or aircraft pilots, and which can provide information redundancy to improve airport security.

Paper Nr: 31
Title:

Artificial Neural Networks for In-silico Experiments on Perception

Authors:

Simon Odense

Abstract: Here the potential use of artificial neural networks for the purpose of understanding the biological processes behind perception is investigated. Current work in computer vision is surveyed focusing on methods to determine how a neural network utilizes it's resources. Analogies between feature detectors in deep neural networks and signaling pathways in the human brain are made. With these analogies in mind, procedures are outlined for experiments on perception using the recurrent temporal restricted Boltzmann machine as an example. The potential use of these experiments to help explain disorders of human perception is then described.

Posters
Paper Nr: 9
Title:

Perceptron Learning for Classification Problems - Impact of Cost-Sensitivity and Outliers Robustness

Authors:

Philippe Thomas

Abstract: In learning approaches for classification problem, the misclassification error types may have different impacts. To take into account this notion of misclassification cost, cost sensitive learning algorithms have been proposed, in particular for the learning of multilayer perceptron. Moreover, data are often corrupted with outliers and in particular with label noise. To respond to this problem, robust criteria have been proposed to reduce the impact of these outliers on the accuracy of the classifier. This paper proposes to associate a cost sensitivity weight to a robust learning rule in order to take into account simultaneously these two problems. The proposed learning rule is tested and compared on a simulation example. The impact of the presence or absence of outliers is investigated. The influence of the costs is also studied. The results show that the using of conjoint cost sensitivity weight and robust criterion allows to improve the classifier accuracy.

Paper Nr: 12
Title:

A Yet Faster Version of Complex-valued Multilayer Perceptron Learning using Singular Regions and Search Pruning

Authors:

Seiya Satoh and Ryohei Nakano

Abstract: In the search space of a complex-valued multilayer perceptron having J hidden units, C-MLP(J), there are singular regions, where the gradient is zero. Although singular regions cause serious stagnation of learning, there exist narrow descending paths from the regions. Based on this observation, a completely new learning method called C-SSF (complex singularity stairs following) 1.0 was proposed, which utilizes singular regions to generate starting points of C-MLP(J) search. Although C-SSF1.0 finds excellent solutions of successive C-MLPs, it takes long CPU time because the number of searches increases as J gets larger. To deal with this problem, C-SSF1.1 was proposed, a few times faster by the introduction of search pruning, but it still remained unsatisfactory. In this paper we propose a yet faster C-SSF1.3, going further with search pruning, and then evaluate the method in terms of solution quality and processing time.

Paper Nr: 32
Title:

Learning-based Distance Evaluation in Robot Vision - A Comparison of ANFIS, MLP, SVR and Bilinear Interpolation Models

Authors:

Hossam Fraihat, Kurosh Madani and Christophe Sabourin

Abstract: This paper deals with visual evaluation of object distances using Soft-Computing based approaches and pseudo-3D standard low-cost sensor, namely the Kinect. The investigated technique points toward robots’ vision and visual metrology of the robot’s surrounding environment. The objective is providing the robot the ability of evaluating distances between objects in its surrounding environment. In fact, although presenting appealing advantages, the Kinect has not been designed for metrological aims. The investigated approach offers the possibility to use this low-cost pseudo-3D sensor for distance evaluation avoiding 3D feature extraction and thus exploiting the simplicity of only 2D image’ processing. Experimental results show the viability of the proposed approach and provide comparison between different machine learning techniques as Adaptive-network-based fuzzy inference (ANFIS), Multi-layer Perceptron (MLP), Support vector regression (SVR), Bilinear interpolation.