NCTA 2016 Abstracts


Full Papers
Paper Nr: 3
Title:

Self-Organizing Maps in the Design of Anti-spam Filters - A Proposal based on Thematic Categories

Authors:

Ylermi Cabrera-León, Patricio García Báez and Carmen Paz Suárez-Araujo

Abstract: Spam, or unsolicited messages sent massively, is one of the threats that affects email and other media. Its high volume generates substantial time and economic losses. A solution to this problem is presented: a hybrid anti-spam filter based on unsupervised Artificial Neural Networks (ANNs). It consists of two steps, preprocessing and processing, both based on different computation models: programmed and neural (using Kohonen SOM). This system has been optimized using, as a data corpus, ham from “Enron Email” and spam from two different sources: traditional (user’s inbox) and spamtrap-honeypot. It has been proved that thematic categories can be found both in spam and ham words. 1260 system configurations were analyzed, comparing their quality and performance with the most used metrics. All of them achieved AUC > 0.90 and the best 204 AUC > 0.95, despite just using 13 attributes for the input vectors of the SOM, one for each thematic category. Results were similar to other researchers’ over the same corpus, though they make use of different Machine Learning (ML) methods and a number of attributes several orders of magnitude greater. It was further tested with datasets not utilized during design, obtaining 0.77 < AUC < 0.96 with normalized data.

Paper Nr: 4
Title:

Noise Resilience of an RGNG-based Grid Cell Model

Authors:

Jochen Kerdels and Gabriele Peters

Abstract: Grid cells are neurons in the entorhinal cortex of mammals that are known for their peculiar, grid-like firing patterns. We developed a generic computational model that describes the behavior of neurons with such firing patterns in terms of a competitive, self-organized learning process. Here we investigate how this process can cope with increasing amounts of noise in its input signal. We demonstrate, that the firing patterns of simulated neurons are mostly unaffected with regard to their structure even if high levels of noise are present in the input. In contrast, the maximum activity of the corresponding neurons decreases significantly with increasing levels of noise. Based on these results we predict that real grid cells can retain their triangular firing patterns in the presence of noise, but may exhibit a noticeable decrease in their peak firing rates.

Paper Nr: 5
Title:

Modelling the Grid-like Encoding of Visual Space in Primates

Authors:

Jochen Kerdels and Gabriele Peters

Abstract: Several regions of the mammalian brain contain neurons that exhibit grid-like firing patterns. The most prominent example of such neurons are grid cells in the entorhinal cortex (EC) whose activity correlates with the animal's location. Correspondingly, contemporary models of grid cells interpret this firing behavior as a specialized, functional part within a system for orientation and navigation. However, Killian et al. report on neurons in the primate EC that show similar, grid-like firing patterns but encode gaze-positions in the field of view instead of locations in the environment. We hypothesized that the phenomenon of grid-like firing patterns may not be restricted to navigational tasks and may be related to a more general, underlying information processing scheme. To explore this idea, we developed a grid cell model based on the recursive growing neural gas (RGNG) algorithm that expresses this notion. Here we show that our grid cell model can -- in contrast to established grid cell models -- also describe the observations of Killian et al. and we outline the general conditions under which we would expect neurons to exhibit grid-like activity patterns in response to input signals independent of a presumed, functional task of the neurons.

Short Papers
Paper Nr: 1
Title:

Determining the Near Optimal Architecture of Autoencoder using Correlation Analysis of the Network Weights

Authors:

Heng Ma, Yonggang Lu and Haitao Zhang

Abstract: Currently, deep learning has already been successfully applied in many fields such as image recognition, recommendation systems and so on. Autoencoder, as an important deep learning model, has attracted a lot of research interests. The performance of the autoencoder can greatly be affected by its architecture. How-ever, how to automatically determine the optimal architecture of the autoencoder is still an open question. Here we propose a novel method for determining the optimal network architecture based on the analysis of the correlation of the network weights. Experiments show that for different datasets the optimal architecture of the autoencoder may be different, and the proposed method can be used to obtain near optimal network architecture separately for different datasets.

Paper Nr: 2
Title:

House Price Estimation from Visual and Textual Features

Authors:

Eman H. Ahmed and Mohamed Moustafa

Abstract: Most existing automatic house price estimation systems rely only on some textual data like its neighborhood area and the number of rooms. The final price is estimated by a human agent who visits the house and assesses it visually. In this paper, we propose extracting visual features from house photographs and combining them with the house’s textual information. The combined features are fed to a fully connected multilayer Neural Network (NN) that estimates the house price as its single output. To train and evaluate our network, we have collected the first houses dataset (to our knowledge) that combines both images and textual attributes. The dataset is composed of 535 sample houses from the state of California, USA. Our experiments showed that adding the visual features increased the R-value by a factor of 3 and decreased the Mean Square Error (MSE) by one order of magnitude compared with textual-only features. Additionally, when trained on the textual-only features housing dataset (Lichman, 2013), our proposed NN still outperformed the existing model published results (Khamis and Kamarudin, 2014).

Paper Nr: 6
Title:

Neural Model for the Influence of Shading on the Multistability of the Perception of Body Motion

Authors:

Leonid Fedorov, Joris Vangeneugden and Martin Giese

Abstract: Body motion perception from impoverished stimuli shows interesting dynamic properties, such as multistability and spontaneous perceptual switching. Psychophysical experiments show that such multistability disappears when the stimulus includes also shading cues along the body surface. Classical neural models for body motion perception have not addressed perceptual multistability. We present an extension of a classical neurodynamic model for biological and body motion perception that accounts for perceptual switching, and its dependence on shading cues on the body surface. We demonstrate that a set of psychophysical observations can be accounted for in a unifying manner by a hierarchical neural model for body motion processing that includes an additional shading pathway, which processes luminance gradients within the individual body segments. The goal of our model is to explain psychophysics and neural mechanism in the brain.

Paper Nr: 8
Title:

Extreme Learning Machine with Enhanced Variation of Activation Functions

Authors:

Jacek Kabzinski

Abstract: The main aim of this paper is to stress the fact that the sufficient variability of activation functions (AF) is important for an Extreme Learning Machine (ELM) approximation accuracy and applicability. A slight modification of the standard ELM procedure is proposed, which allows increasing the variance of each AF, without losing too much from the simplicity of random selection of parameters. The proposed modification does not increase the computational complexity of an ELM training significantly. Enhancing the variation of AFs results in reduced output weights norm, better numerical conditioning of the output weights calculation, smaller errors for the same number of the hidden neurons. The proposed approach works efficiently together with the Tikhonov regularization of ELM.

Paper Nr: 9
Title:

Exploring the Neurosolver in Playing Adversarial Games

Authors:

Andrzej Bieszczad

Abstract: In the past, the Neurosolver, a neuromorphic planner and a general problem solver, was used in several exploratory applications, such as Blocks World and Towers of Hanoi puzzles, that in which we investigated its problem solving capabilities. In all of them, there was only one agent that had a single point of view focus: how to solve a posed problem by generating a sequence of actions to get the system from its current state to some goal state. In this paper, we report on our experiments with exploring the Neurosolver’s capabilities to deal with more sophisticated challenges in solving problems. For that purpose, we employed the Neurosolver as a driver for adversary games. In that kind of environment, the Neurosolver cannot just generate a plan and then follow it through. Instead, the plan has to be revised dynamically step by step in response to the other actors following their own plans realizing adversarial points of view. We conclude that while the Neurosolver can learn to play an adversarial game, to play it well it would need a good teacher.

Posters
Paper Nr: 11
Title:

Theoretical Notes on Unsupervised Learning in Deep Neural Networks

Authors:

Vladimir Golovko and Aliaksandr Kroshchanka

Abstract: Over the last decade the deep neural networks are the powerful tool in the domain of machine learning. The important problem is training of deep neural network, because learning of such a network is much complicated compared to shallow neural networks. This is due to the vanishing gradient problem, poor local minima and unstable gradient problem. Therefore a lot of deep learning techniques were developed that permit us to overcome some limitations of conventional training approaches. In this paper we investigate the unsupervised learning in deep neural networks. We have proved that maximization of the log-likelihood input data distribution of restricted Boltzmann machine is equivalent to minimizing the cross-entropy and to special case of minimizing the mean squared error. The main contribution of this paper is a novel view and new understanding of an unsupervised learning in deep neural networks.

Paper Nr: 12
Title:

Beyond Machine Learning: Autonomous Learning

Authors:

Frédéric Alexandre

Abstract: Recently, Machine Learning has achieved impressive results, surpassing human performances, but these powerful algorithms are still unable to define their goals by themselves or to adapt when the task changes. In short, they are not autonomous. In this paper, we explain why autonomy is an important criterion for really powerful learning algorithms. We propose a number of characteristics that make humans more autonomous than machines when they learn. Humans have a system of memories where one memory can compensate or train another memory if needed. They are able to detect uncertainties and adapt accordingly. They are able to define their goals by themselves, from internal and external cues and are capable of self-evaluation to adapt their learning behavior. We also suggest that introducing these characteristics in the domain of Machine Learning is a critical challenge for future intelligent systems.