NCTA 2020 Abstracts


Full Papers
Paper Nr: 5
Title:

InfoSalGAIL: Visual Attention-empowered Imitation Learning of Pedestrian Behavior in Critical Traffic Scenarios

Authors:

Igor Vozniak, Matthias Klusch, André Antakli and Christian Müller

Abstract: The imitation learning of complex pedestrian behavior based on visual input is a challenge due to the underlying large state space and variations. In this paper, we present a novel visual attention-based imitation learning framework, named InfoSalGAIL, for end-to-end imitation learning of (safe, unsafe) pedestrian navigation policies through visual expert demonstrations empowered by eye fixation sequence and augmented reward function. This work shows the relation in latent space between the policy estimated trajectories and visual-attention map. Moreover, the conducted experiments revealed that InfoSalGAIL can significantly outperform the state-of-the-art baseline InfoGAIL. In fact, its visual attention-empowered imitation learning tends to much better generalize the overall policy of pedestrian behavior leveraging apprenticeship learning to generate more human-like pedestrian trajectories in virtual traffic scenes with the open source driving simulator OpenDS. InfoSalGAIL can be utilized in the process of generating and validating critical scenarios for adaptive driving assistance systems.

Paper Nr: 12
Title:

Complexity vs. Performance in Granular Embedding Spaces for Graph Classification

Authors:

Luca Baldini, Alessio Martino and Antonello Rizzi

Abstract: The most distinctive trait in structural pattern recognition in graph domain is the ability to deal with the organization and relations between the constituent entities of the pattern. Even if this can be convenient and/or necessary in many contexts, most of the state-of the art classification techniques can not be deployed directly in the graph domain without first embedding graph patterns towards a metric space. Granular Computing is a powerful information processing paradigm that can be employed in order to drive the synthesis of automatic embedding spaces from structured domains. In this paper we investigate several classification techniques starting from Granular Computing-based embedding procedures and provide a thorough overview in terms of model complexity, embedding space complexity and performances on several open-access datasets for graph classification. We witness that certain classification techniques perform poorly both from the point of view of complexity and learning performances as the case of non-linear SVM, suggesting that high dimensionality of the synthesized embedding space can negatively affect the effectiveness of these approaches. On the other hand, linear support vector machines, neuro-fuzzy networks and nearest neighbour classifiers have comparable performances in terms of accuracy, with second being the most competitive in terms of structural complexity and the latter being the most competitive in terms of embedding space dimensionality.

Paper Nr: 14
Title:

Mining M-Grams by a Granular Computing Approach for Text Classification

Authors:

Antonino Capillo, Enrico de Santis, Fabio F. Mascioli and Antonello Rizzi

Abstract: Text mining and text classification are gaining more and more importance in AI related research fields. Researchers are particularly focused on classification systems, based on structured data (such as sequences or graphs), facing the challenge of synthesizing interpretable models, exploiting gray-box approaches. In this paper, a novel gray-box text classifier is presented. Documents to be classified are split into their constituent words, or tokens. Groups of frequent m tokens (or m-grams) are suitably mined adopting the Granular Computing framework. By fastText algorithm, each token is encoded in a real-valued vector and a custom-based dissimilarity measure, grounded on the Edit family, is designed specifically to deal with m-grams. Through a clustering procedure the most representative m-grams, pertaining the corpus of documents, are extrapolated and arranged into a Symbolic Histogram representation. The latter allows embedding documents in a well-suited real-valued space in which a standard classifier, such as SVM, can safety operate. Along with the classification procedure, an Evolutionary Algorithm is in charge of performing features selection, which is able to select most relevant symbols – m-grams – for each class. This study shows how symbols can be fruitfully interpreted, allowing an interesting knowledge discovery procedure, in lights with the new requirements of modern explainable AI systems. The effectiveness of the proposed algorithm has been proved through a set of experiments on paper abstracts classification and SMS spam detection.

Paper Nr: 18
Title:

Improving the Training of Convolutional Neural Network using Between-class Distance

Authors:

Jiani Liu, Xiang Zhang and Yonggang Lu

Abstract: Recently, Convolutional Neural Networks (CNN) have demonstrated state-of-the-art image classification performance. However, in many cases, it is hard to train the network optimally in multi-class classification. One way to alleviate the problem is to make good use of the training data, and more research work needs to be done on how to use the training data in multi-class classification more efficiently. In this paper we propose a method to make the classification more accurate by analyzing the between-class distance of the deep features of the training data. The specific pattern of the between-class distances is used to improve the training process. It is shown that the proposed method can improve the training on both MNIST and EMNIST datasets.

Paper Nr: 19
Title:

Batch Contrastive Regularization for Deep Neural Network

Authors:

Muhammad Tanveer, Hung K. Tan, Hui F. Ng, Maylor K. Leung and Joon H. Chuah

Abstract: As neural network becomes deeper, it becomes more capable of generating more powerful representation for a wide variety of tasks. However, deep neural network has a large number of parameters and easy to overfit the training samples. In this paper, we present a new regularization technique, called batch contrastive regularization. Regularization is performed by comparing samples collectively via contrastive loss which encourages intra-class compactness and inter-class separability in an embedded Euclidean space. To facilitate learning of embedding features for contrastive loss, a two-headed neural network architecture is used to decouple regularization classification. During inference, the regularization head is discarded and the network operates like any conventional classification network. We also introduce bag sampling to ensure sufficient positive samples for the classes in each batch. The performance of the proposed architecture is evaluated on CIFAR-10 and CIFAR-100 databases. Our experiments show that features regularized by contrastive loss has strong generalization performance, yielding over 8% improvement on ResNet50 for CIFAR-100 when trained from scratch.

Short Papers
Paper Nr: 3
Title:

Reflexive Reinforcement Learning: Methods for Self-Referential Autonomous Learning

Authors:

B. I. Lyons and J. M. Herrmann

Abstract: Reinforcement learning aims at maximising an external evaluative signal over a certain time horizon. If no reward is available within the time horizon, the agent faces an autonomous learning task which can be used to explore, to gather information, and to bootstrap particular learning behaviours. We discuss here how the agent can use a current representation of the value, of its state and of the environment, in order to produce autonomous learning behaviour in the absence of a meaningful rewards. The family of methods that is introduced here is open to further development and research in the field of reflexive reinforcement learning.

Paper Nr: 8
Title:

On the Robustness of Convolutional Neural Networks Regarding Transformed Input Images

Authors:

Frederik Timme, Jochen Kerdels and Gabriele Peters

Abstract: Convolutional Neural Networks (CNNs) have become the dominant and arguably most successful approach for the task of image classification since the release of AlexNet in 2012. Despite their excellent performance, CNNs continue to suffer from a still poorly understood lack of robustness when confronted with adversarial attacks or particular forms of handcrafted datasets. Here we investigate how the recognition performance of three widely used CNN architectures (AlexNet, VGG19 and ResNeXt) changes in response to certain input data transformations. 10,000 images from the ILSVRC2012s validation dataset were systematically manipulated by means of common transformations (translation, rotation, color change, background replacement) as well as methods like image collages and jigsaw-like puzzles. Both the effect of single and combined transformations are investigated. Our results show that three of these input image manipulations (rotation, collage, and puzzle) can cause a significant drop in classification accuracy in all evaluated architectures. In general, the more recent VGG19 and ResNeXt displayed a higher robustness than AlexNet in our experiments indicating that some progress has been made to harden the CNN approach against malicious or unforeseen input.

Paper Nr: 9
Title:

Bootstrapping a DQN Replay Memory with Synthetic Experiences

Authors:

Wenzel P. von Pilchau, Anthony Stein and Jörg Hähner

Abstract: An important component of many Deep Reinforcement Learning algorithms is the Experience Replay that serves as a storage mechanism or memory of experienced transitions. These experiences are used for training and help the agent to stably find the perfect trajectory through the problem space. The classic Experience Replay however makes only use of the experiences it actually made, but the stored transitions bear great potential in form of knowledge about the problem that can be extracted. The gathered knowledge contains state-transitions and received rewards that can be utilized to approximate a model of the environment. We present an algorithm that creates synthetic experiences in a nondeterministic discrete environment to assist the learner with augmented training data. The Interpolated Experience Replay is evaluated on the FrozenLake environment and we show that it can achieve a 17% increased mean reward compared to the classic version.

Paper Nr: 13
Title:

Intrusion Detection in Wi-Fi Networks by Modular and Optimized Ensemble of Classifiers

Authors:

Giuseppe Granato, Alessio Martino, Luca Baldini and Antonello Rizzi

Abstract: With the breakthrough of pervasive advanced networking infrastructures and paradigms such as 5G and IoT, cybersecurity became an active and crucial field in the last years. Furthermore, machine learning techniques are gaining more and more attention as prospective tools for mining of (possibly malicious) packet traces and automatic synthesis of network intrusion detection systems. In this work, we propose a modular ensemble of classifiers for spotting malicious attacks on Wi-Fi networks. Each classifier in the ensemble is tailored to characterize a given attack class and is individually optimized by means of a genetic algorithm wrapper with the dual goal of hyper-parameters tuning and retaining only relevant features for a specific attack class. Our approach also considers a novel false alarm management procedure thanks to a proper reliability measure formulation. The proposed system has been tested on the well-known AWID dataset, showing performances comparable with other state of the art works both in terms of accuracy and knowledge discovery capabilities. Our system is also characterized by a modular design of the classification model, allowing to include new possible attack classes in an efficient way.

Paper Nr: 16
Title:

Risk Estimation in Data-driven Fault Prediction for a Biomass-fired Power Plant

Authors:

Ivan Ryzhikov, Mika Liukkonen, Ari Kettunen and Yrjö Hiltunen

Abstract: In this study, we consider a fault prediction problem for the case when there are no variables by which we could determine that the system is in the fault state. We propose an approach that is based on constructing auxiliary variable, thus it is possible to reduce the initial problem to the supervised learning problem of risk estimation. The suggested target variable is an indicator showing how close the system is to the fault that is why we call it a risk estimation variable. The risk is growing some time before the actual fault has happened and reaches the highest value in that timestamp, but there is a high level of uncertainty for the times when the system has been operating normally. We suggest specific criterion that takes uncertainty of risk estimation into account by tuning three weighting coefficients. Finally, the supervised learning problem with risk variable and specific criterion can be solved by the means of machine learning. This work confirm that data-driven risk estimation can be integrated into digital services to successfully manage plant operational changes and support plant prescriptive maintenance. This was demonstrated with data from a commercial circulating fluidized bed firing various biomass and residues but is generally applicable to other production plants.

Paper Nr: 17
Title:

Empirical Evaluation of Distance Measures for Nearest Point with Indexing Ratio Clustering Algorithm

Authors:

Raneem Qaddoura, Hossam Faris, Ibrahim Aljarah, J. J. Merelo and Pedro A. Castillo

Abstract: Selecting the proper distance measure is very challenging for most clustering algorithms. Some common distance measures include Manhattan (City-block), Euclidean, Minkowski, and Chebyshev. The so called Nearest Point with Indexing Ratio (NPIR) is a recent clustering algorithm, which tries to overcome the limitations of other algorithms by identifying arbitrary shapes of clusters, non-spherical distribution of points, and shapes with different densities. It does so by iteratively utilizing the nearest neighbors search technique to find different clusters. The current implementation of the algorithm considers the Euclidean distance measure, which is used for the experiments presented in the original paper of the algorithm. In this paper, the impact of the four common distance measures on NPIR clustering algorithm is investigated. The performance of NPIR algorithm in accordance to purity and entropy measures is investigated on nine data sets. The comparative study demonstrates that the NPIR generates better results when Manhattan distance measure is used compared to the other distance measures for the studied high dimensional data sets in terms of purity and entropy.

Paper Nr: 20
Title:

Classification of Alzheimer’s Disease, Mild Cognitive Impairment, and Normal Controls with Multilayer Perceptron Neural Network and Neuropsychological Test Data

Authors:

Ibrahim Almubark, Samah Alsegehy, Xiong Jiang and Lin-Ching Chang

Abstract: Recent advances in machine learning have shown outstanding performances in biological and medical data analysis to assist for early detection, diagnosis, and treatment of diseases. Alzheimer's disease (AD) is a neurodegenerative disease and the most common cause of dementia in older adults. In this study, multilayer perceptron (MLP) neural networks are developed to classify AD, Mild Cognitive Impairment (MCI), and Cognitive Normal (CN) subjects based upon the data from standard neuropsychological tests. Three neuropsychological tests from the Alzheimer’s Disease Neuroimaging Initiative (ADNI) database, Alzheimer’s Disease Assessment Scale-Cognitive subscale (ADAS-Cog), Mini-Mental State Examination (MMSE), and Functional Activities Questionnaire (FAQ), were used to train MLP neural networks. We first build three MLP models that can classify AD vs. CN, AD vs. MCI, and MCI vs. CN. We then construct a 3-way MLP classifier to classify AD vs. MCI vs. CN. Finally, we propose a cascade 3-way classification method to further improve the model performance. Using the neuropsychological test data from ADNI database, our result shows the pairwise MLP models (i.e., AD vs. CN, AD vs. MCI, and MCI vs. CN) have the accuracy of 99.760.48, 89.643.94, and 90.812.91, respectively. The multi-class MLP model has an average accuracy of 84.283.66, and the proposed cascaded MLP approach further improves the performance of the multi-class classification with an average accuracy of 86.263.15.

Paper Nr: 21
Title:

Neural Semantic Pointers in Context

Authors:

Alessio Plebe and Arianna Pavone

Abstract: Resolving linguistic ambiguities is a task frequently called for in human communication. In many cases, such task cannot be solved without additional information about an associated context, which can be often captured from the visual scene referred by the sentence. This type of inference is crucial in several aspects of language, communication in the first place, and in the grounding of language in perception. This paper focuses on the contextual effects of visual scenes on semantics, investigated using neural computational simulation. Specifically, here we address the problem of selecting the interpretation of sentences with an ambiguous prepositional phrase, matching the context provided by visual perception. More formally, provided with a sentence, admitting two or more candidate resolutions for a prepositional phrase attachment, and an image that depicts the content of the sentence, it is required to choose the correct resolution depending on the image’s content. From the neuro-computational point of view, our model is based on Nengo, the implementation of Neural Engineering Framework (NEF), whose basic semantic component is the so-called Semantic Pointer Architecture (SPA), a biologically plausible way of representing concepts by dynamic neural assemblies. We evaluated the ability of our model in resolving linguistic ambiguities on the LAVA (Language and Vision Ambiguities) dataset, a corpus of sentences with a wide range of ambiguities, associated with visual scenes.

Paper Nr: 22
Title:

Classifying Incomplete Vectors using Decision Trees

Authors:

Bhekisipho Twala, Raj Pillay and Ramapulana Nkoana

Abstract: An attempt is made to address the problem of classifying incomplete vectors using decision trees. The essence of the approach is the proposal that in supervised learning classification of incomplete vectors can be improved in probabilistic terms. This approach, which is based on the a priori probability of each value determined from the instances at that node of the tree that has specified values, first exploits the total probability and Bayes’ theorems and then the probit and logit model probabilities. The proposed approach (developed in three versions) is evaluated using 21 machine learning datasets from its effect or tolerance of incomplete test data. Experimental results are reported, showing the effectiveness of the proposed approach in comparison with multiple imputation and fractioning of instances strategy.

Paper Nr: 25
Title:

Generative Modeling of Synthetic Eye-tracking Data: NLP-based Approach with Recurrent Neural Networks

Authors:

Mahmoud Elbattah, Jean-Luc Guérin, Romuald Carette, Federica Cilia and Gilles Dequen

Abstract: This study explores a Machine Learning-based approach for generating synthetic eye-tracking data. In this respect, a novel application of Recurrent Neural Networks is experimented. Our approach is based on learning the sequence patterns of eye-tracking data. The key idea is to represent eye-tracking records as textual strings, which describe the sequences of fixations and saccades. The study therefore could borrow methods from the Natural Language Processing (NLP) domain for transforming the raw eye-tracking data. The NLP-based transformation is utilised to convert the high-dimensional eye-tracking data into an amenable representation for learning. Furthermore, the generative modeling could be implemented as a task of text generation. Our empirical experiments support further exploration and development of such NLP-driven approaches for the purpose of producing synthetic eye-tracking datasets for a variety of potential applications.

Paper Nr: 7
Title:

Soil Moisture Prediction Model from ERA5-Land Parameters using a Deep Neural Networks

Authors:

Daouda Diouf, Carlos Mejia and Djibril Seck

Abstract: In a global context of scarcity of water resources, accurate prediction of soil moisture is important for its rational use and management. Soil moisture is included in the list of Essential Climate Variables. Because of the complex soil structure, meteorological parameters and the diversity of vegetation cover, it is not easy to establish a predictive relationship of soil moisture. In this paper, using the large amounts of data obtained in West Africa, we set up a deep neural network to establish an estimation of soil moisture for the two first layers and its prediction temporally and spatially. We construct deep neural network model which predicts soil moisture layer 1 and layer 2 multiple days in the future. Results obtained for accuracy training and test are greater than 93 %. The mean absolute errors are very low and vary between 0,01 to 0,03 m3/m3.

Paper Nr: 23
Title:

Learning Spatial Relations with a Standard Convolutional Neural Network

Authors:

Kevin Swingler and Mandy Bath

Abstract: This paper shows how a standard convolutional neural network (CNN) without recurrent connections is able to learn general spatial relationships between different objects in an image. A dataset was constructed by placing objects from the Fashion-MNIST dataset onto a larger canvas in various relational locations (for example, trousers left of a shirt, both above a bag). CNNs were trained to name the objects and their spatial relationship. Models were trained to perform two different types of task. The first was to name the objects and their relationships and the second was to answer relational questions such as “Where is the shoe in relation to the bag?”. The models performed at above 80% accuracy on test data. The models were also capable of generalising to spatial combinations that had been intentionally excluded from the training data.

Paper Nr: 24
Title:

Computation of Neural Networks Lyapunov Functions for Discrete and Continuous Time Systems with Domain of Attraction Maximization

Authors:

Benjamin Bocquillon, Philippe Feyel, Guillaume Sandou and Pedro Rodriguez-Ayerbe

Abstract: This contribution deals with a new approach for computing Lyapunov functions represented by neural networks for nonlinear discrete-time systems to prove asymptotic stability. Based on the Lyapunov theory and the notion of domain of attraction, the proposed approach deals with an optimization method for determining a Lyapunov function modeled by a neural network while maximizing the domain of attraction. Several simulation examples are presented to illustrate the potential of the proposed method.