Registered Data

[CT123]

[00664] A New Sampling Technique for Learning with Hypergraph Neural Networks

  • Session Date & Time : 3C (Aug.23, 13:20-15:00)
  • Type : Contributed Talk
  • Abstract : Hypergraphs can represent higher-order relations among objects. Traditional hypergraph neural networks produce high computational cost and timing. We propose a new sampling technique for learning with hypergraph neural networks. The core idea is to design a layer-wise sampling scheme for nodes and hyperedges to approximate original hypergraph convolution. Notably, the proposed sampling technique allows us to handle large-scale hypergraph learning. Experiment results demonstrate that our proposed model keeps a good balance between time and accuracy.
  • Classification : 68T07, 05C65, 62D05, 68T09, large-scale hypergraph learning, hypergraph neural networks, hypergraph sampling, variance reduction, importance sampling
  • Author(s) :
    • Fengcheng Lu (The University of Hong Kong)
    • Michael Kwok-Po Ng (The University of Hong Kong)

[00773] Machine Learning Model for Thin Metal Sheet Counting and Thickness Measurement

  • Session Date & Time : 3C (Aug.23, 13:20-15:00)
  • Type : Contributed Talk
  • Abstract : In this talk we are to discuss about counting stacked metal foils in real time. For the purpose a non-contact method based on broadband X-ray absorption spectra was employed to scan the experimental samples and artificial neural network was built to count and measure thickness of the stacked foil. Further, the attained results are compared with polynomial fitting model and linear regression in order to verify the difference in prediction accuracy of the three models.
  • Classification : 68T07, 78M32, 82C32, 62J05, 65D10
  • Author(s) :
    • Elayaperumal Ayyasamy (Anna University, Chennai)

[00817] Understanding Flood Flow Physics via Data-Informed Learning

  • Session Date & Time : 3C (Aug.23, 13:20-15:00)
  • Type : Contributed Talk
  • Abstract : Modeling the dynamics of fast-moving floods has historically been an intractable problem due to the inherent complexity and multi-scale physics of the underlying processes involved. Recent advancements in physics-constrained machine learning indicate that neural networks can be used to effectively model phenomena for which physical laws are poorly understood. By combining real data and first principles, we show that we can enhance knowledge about the underlying physics of flood phenomena via the learned constitutive laws.
  • Classification : 68T07, 76T99, 86-10
  • Author(s) :
    • Jonathan Thompson (University of Colorado Colorado Springs)
    • Radu Cascaval (University of Colorado Colorado Springs)

[00856] Successive image generation though cyclic transformations using CycleGAN

  • Session Date & Time : 3C (Aug.23, 13:20-15:00)
  • Type : Contributed Talk
  • Abstract : CycleGAN is a deep generative adversarial networks that performs image to image style translation by learning relationship between two image domains. By using CycleGAN, here we developed a model performing cyclical transformation that generates a series of similar images. This system can be regarded as a dynamical system; it can continuously sample various images along the trajectory of the dynamical system. The chaotic behavior of this deep model was studied.
  • Classification : 68T07, 37N99
  • Author(s) :
    • Takaya Tanaka (Graduate School of Engineering, Fukuoka Institute of Technology)
    • Takaya Tanaka (Graduate School of Engineering, Fukuoka Institute of Technology)
    • Yutaka Yamaguti (Faculty of Information Engineering, Fukuoka Institute of Technology)

[01057] Hypergraph Attention Graph Convolution Network for Precision Medicine

  • Session Date & Time : 3C (Aug.23, 13:20-15:00)
  • Type : Industrial Contributed Talk
  • Abstract : Graph neural networks have achieved prominent results recently in many research fields. However, the existing GCN approaches encountered large challenges when handling some issues with heterogeneous graphs of a huge number of nodes, and multiple types of relations. A good example of such issues can be the relational networks of adverse events and drugs from the area of precision medicine. We propose the Hypergraph Attention convolution mechanism as a promising remedy and is worth further studies.
  • Classification : 68T07, 68T30
  • Author(s) :
    • ZHOUMING XU (The University of Hong Kong)

[01196] Deep Solvers in Shape Optimization

  • Session Date & Time : 3D (Aug.23, 15:30-17:10)
  • Type : Contributed Talk
  • Abstract : We introduce a novel mesh-free method for computing the shape derivative in PDE-constrained shape optimization problems. Our approach is based on a probabilistic deep solver, which can be shown to converge for a wide class of seminilinear PDEs, and a suitable representation of the shape gradient. In contrast to finite element, volume and difference methods, our approach does not require a discretization of the domain’s interior. We also present examples for performance illustration.
  • Classification : 68T07, 65N99
  • Author(s) :
    • Maximilian Würschmidt (Trier University)
    • Frank Seifried (Trier University)
    • Luka Schlegel (Trier University)
    • Volker Schulz (Trier University)

[01477] Optimization of a submerged piezoelectric device using an ANN Model

  • Session Date & Time : 3D (Aug.23, 15:30-17:10)
  • Type : Contributed Talk
  • Abstract : The design of a submerged piezoelectric wave energy converter (PWEC) device has been analyzed to optimize the power generated by the PWEC device. An artificial neural network (ANN) is adopted to optimize the geometric parameters of the device. First, a numerical model is introduced using the boundary element methodology (BEM). The input database for the modeling of the ANN model is generated using the Latin Hypercube Sampling method, and the output database for the modeling of the ANN model is simulated using the numerical model based on BEM. Four hundred samples are used to model the ANN with data taken in a 70:30 ratio for training and validation of the model. The prediction of the optimal parameter values for the design of the PWEC device is carried out using a database containing 3000 sample points generated randomly using the LHS method. The developed ANN model shows a good agreement between the training accuracy and the validation accuracy. Also, the model forecast provides a range for the geometric parameters of the PWEC device to optimize power generation.
  • Classification : 68T07, 68T20, 68V99
  • Author(s) :
    • Vipin V (Birla Institute of Technology and Science Pilani, Hyderabad Campus)
    • SANTANU KOLEY (Dept.of Mathematics, Birla Institute of Technology and Science - Pilani, Hyderabad Campus)

[01500] Lower bounds on the Generalization Error of Nonlinear Learning Models in High Dimensions

  • Session Date & Time : 3D (Aug.23, 15:30-17:10)
  • Type : Contributed Talk
  • Abstract : Modern learning algorithms often contain more parameters than training samples. Despite their huge complexity, their generalization error is small. In this talk, we study generalization properties of algorithms in high dimensions. We provide lower bounds on the generalization error in various settings for any algorithm, in particular for deep neural network. The bounds are calculated using random matrix theory. We will explore different regimes of overparameterization and show in which cases such bounds are tight.
  • Classification : 68T07, 62F15, 62M45, 62F12, 60B20
  • Author(s) :
    • Inbar Seroussi (Tei-Aviv University)

[01817] Time-series medical data classification using echo state network

  • Session Date & Time : 3D (Aug.23, 15:30-17:10)
  • Type : Contributed Talk
  • Abstract : Most time-series medical data classification tasks are carried out using deep recurrent neural networks. However, deep neural networks tend to consume enormous computational power. Echo state network is an efficient model for processing temporal data due to its low training cost. The reservoir maps input signals into a high-dimensional dynamical system and the readout layer extracts patterns from it. Therefore, we developed a new methodology that can classify time-series data using echo state network.
  • Classification : 68T07, 62R07, 62P10
  • Author(s) :
    • Zonglun Li (University College London)

[02164] Uncertainty-Aware Null Space Networks for Data-Consistent Image Reconstruction

  • Session Date & Time : 3D (Aug.23, 15:30-17:10)
  • Type : Contributed Talk
  • Abstract : State-of-the-art reconstruction methods in inverse problems have been developed by incorporating latest advances in deep learning. Before learning approaches can be used in safety-critical areas like medical imaging, a model must not only provide a reconstruction, but also an estimate of its reliability. This study presents a cascaded architecture of null space networks and combines it with recent progress of uncertainty quantification in computer vision. This way, two key properties are met: data-consistency and uncertainty-awareness.
  • Classification : 68T07, 68T37, 92C50, 92C55
  • Author(s) :
    • Christoph Angermann (VASCage – Research Centre on Vascular Ageing and Stroke)
    • Simon Goeppel (Universität Innsbruck)
    • Markus Haltmeier (Universität Innsbruck)

[02276] Detection Topic of Bjorka Using LSTM with LDA

  • Session Date & Time : 3E (Aug.23, 17:40-19:20)
  • Type : Contributed Talk
  • Abstract : This paper presents topic modeling for Hacker Bjorka using LSTM with LDA. The Research purpose is to analyze the public opinion of Bjorka and the topics related to him in the online community. The findings reveal that the majority of the public perception of Bjorka is positive, with accuracy of 80,26% and perplexity of -8,28. This study provides a valuable contribution to the field of computational text analysis and its applications in the online community.
  • Classification : 68T07
  • Author(s) :
    • Muhammad Muhajir (Universitas Gadjah Mada )
    • Dedi Rosadi (Universitas Gadjah Mada )

[02326] Multi-level Wavelet Convolutional Neural Networks for Classifying Lung Cancer

  • Session Date & Time : 3E (Aug.23, 17:40-19:20)
  • Type : Contributed Talk
  • Abstract : Lung cancer classification becomes significant as it can increase the survival rate. Our previous study to classify lung cancer using Recurrent Neural Networks (RNN) and Wavelet RNN intensify accuracy by 2.7%. It may increase with the computational complexity. Thus, in this research, we focus classifying lung cancer using deep learning Neural Networks that has been successfully applied in practice. A model called Multi-level Wavelet Convolutional Neural Networks. This study provides more discussion on Neural Networks and lung cancer classification.
  • Classification : 68T07
  • Author(s) :
    • Devi Nurtiyasari (Gadjah Mada University)
    • Dedi Rosadi (Gadjah Mada University)
    • Abdurakhman Abdurakhman (Gadjah Mada University)

[02329] Improve Error Prediction Using Regularization Model for Movie Recommendation System

  • Session Date & Time : 3E (Aug.23, 17:40-19:20)
  • Type : Contributed Talk
  • Abstract : Currently, most applications (such as Netflix, Spotify, and the others) provide engaging facilities to improve the user’s experience. These applications highly depend on the effectiveness of their recommendation systems. The goal for this paper was to improve error prediction (RMSE and MAE) using Regularization model compared with state-of-art models. The proposed technique obtains a better result than a state-of-art model with an improvement of 0.48% and 1.43% on error prediction using ML-1M dataset, respectively.
  • Classification : 68T07, 68T09, Machine Learning
  • Author(s) :
    • Malim Muhammad (Universitas Gadjah Mada)
    • Dedi Rosadi (Universitas Gadjah Mada)
    • Danardono Danardono (Universitas Gadjah Mada)

[00180] Relationship between musical notes and socio-political events

  • Session Date & Time : 3E (Aug.23, 17:40-19:20)
  • Type : Contributed Talk
  • Abstract : Historians and scientists long suspected that sounds and music impact different cultures. However, empirical data to support such claim is sparse. Previous research using Supervised Machine Learning algorithms, i.e. ANFIS (Adaptive Neuro-Fuzzy Inference System) has successfully categorised musical genre classification and predicted the outcome of the United Kingdom's election results using popular music released in that period by feeding sound wave features to the ANFIS algorithm. This study reports similar research for the Moroccan elections using two different supervised machine learning algorithms namely, k-NN and SVM.
  • Classification : 68T09, 91C99
  • Author(s) :
    • Choi-Hong Lai (University of Greenwich)
    • Nakunam Kokulan (University of Greenwich)
    • Yahya Chahine (University of Greenwich)

[01150] Optimizing Tool Assignment Using Smart Lockers

  • Session Date & Time : 3E (Aug.23, 17:40-19:20)
  • Type : Industrial Contributed Talk
  • Abstract : A logistics operator adopted smart lockers to distribute 150 commercial tools such as tablets and scanners to employees, working in 3 shifts. The lockers gather significant data like tools deficiencies, breakdowns, usage time, punctuality. The assignment policy to deliver tools must minimize downtime using two figure of merit, one classifying employees according to their ability to handle tools, and another measuring the frequency of deficiencies and breakdowns. The proposal is inspired in genetic algorithms.
  • Classification : 68T20, 90B06
  • Author(s) :
    • Jose Alberto Fonseca (Instituto de Telecomunicações - Universidade de Aveiro)
    • Joaquim Ferreira (Instituto de Telecomunicacoes, Universidade de Aveiro)
    • Ricardo Bandeira (Microio,Lda)
    • Fernanda Coutinho (Instituto Superior de Engenharia de Coimbra - IPP)