NCTA 2022 Abstracts

Full Papers Short Papers
Full Papers
Paper Nr: 2
Title:

Interpolated Experience Replay for Continuous Environments
Authors:

Wenzel P. von Pilchau, Anthony Stein and Jörg Hähner
Abstract: The concept of Experience Replay is a crucial element in Deep Reinforcement Learning algorithms of the DQN family. The basic approach reuses stored experiences to, amongst other reasons, overcome the problem of catastrophic forgetting and as a result stabilize learning. However, only experiences that the learner observed in the past are used for updates. We anticipate that these experiences posses additional valuable information about the underlying problem that just needs to be extracted in the right way. To achieve this, we present the Interpolated Experience Replay technique that leverages stored experiences to create new, synthetic ones by means of interpolation. A previous proposed concept for discrete-state environments is extended to work in continuous problem spaces. We evaluate our approach on the MountainCar benchmark environment and demonstrate its promising potential.
Download
Paper Nr: 7
Title:

Explaining Reject Options of Learning Vector Quantization Classifiers
Authors:

André Artelt, Johannes Brinkrolf, Roel Visser and Barbara Hammer
Abstract: While machine learning models are usually assumed to always output a prediction, there also exist extensions in the form of reject options which allow the model to reject inputs where only a prediction with an unacceptably low certainty would be possible. With the ongoing rise of eXplainable AI, a lot of methods for explaining model predictions have been developed. However, understanding why a given input was rejected, instead of being classified by the model, is also of interest. Surprisingly, explanations of rejects have not been considered so far. We propose to use counterfactual explanations for explaining rejects and investigate how to efficiently compute counterfactual explanations of different reject options for an important class of models, namely prototype-based classifiers such as learning vector quantization models.
Download
Paper Nr: 8
Title:

A Suite of Incremental Image Degradation Operators for Testing Image Classification Algorithms
Authors:

Kevin Swingler
Abstract: Convolutional Neural Networks (CNN) are extremely popular for modelling sound and images, but they suffer from a lack of robustness that could threaten their usefulness in applications where reliability is important. Recent studies have shown how it is possible to maliciously create adversarial images—those that appear to the human observer as perfect examples of one class but that fool a CNN into assigning them to a different, incorrect class. It takes some effort to make these images as they need to be designed specifically to fool a given network. In this paper we show that images can be degraded in a number of simple ways that do not need careful design and that would not affect the ability of a human observer, but which cause severe deterioration in the performance of three different CNN models. We call the speed of the deterioration in performance due to incremental degradations in image quality the degradation profile of a model and argue that reporting the degradation profile is as important as reporting performance on clean images.
Download
Paper Nr: 9
Title:

Keep It Simple: Local Search-based Latent Space Editing
Authors:

Andreas Meißner, Andreas Fröhlich and Michaela Geierhos
Abstract: Semantic image editing allows users to selectively change entire image attributes in a controlled manner with just a few clicks. Most approaches use a generative adversarial network (GAN) for this task to learn an appropriate latent space representation and attribute-specific transformations. While earlier approaches often suffer from entangled attribute manipulations, newer ones improve on this aspect by using separate specialized networks for attribute extraction. Iterative optimization algorithms based on backpropagation constitute a possible approach to find attribute vectors with little entanglement. However, this requires a large amount of GPU memory, training instabilities can occur, and the used models have to be differentiable. To address these issues, we propose a local search-based approach for latent space editing. We show that it performs at the same level as previous algorithms and avoids these drawbacks.
Download
Paper Nr: 12
Title:

Spatial Simulation of the Müller-Lyer Illusion Genesis with Convolutional Neural Networks
Authors:

Anton N. Mamaev and Ivan A. Gorbunov
Abstract: The Müller-Lyer illusion is a well-known optical phenomenon with several competing explanations. In the current study we reviewed the illusion in a convolutional neural network from a perspective of image-source relationships in the process of visual functions development. To recreate the effect of the illusion we proposed a novel method that lets us simulate the development of visual functions in a controlled spatial environment from the state of ‘blank slate’ to effective spatial problem solving. This process is designed to reflect the development of human visual system and enable us to determine how depth perception can contribute to the appearance of the phenomenon. We were able to successfully reproduce the effect of the classic Müller-Lyer in 30 independent convolutional models and also get similar results with the variants of the illusion that are thought to be unrelated to spatial perception. For the pairs of classic stimuli we conducted additional statistical analysis using both frequentist and Bayesian methods. The methodological and empirical insights of this study may be helpful for subsequent investigation of visual cognition and reconsideration of the image-source relationships in optical illusions.
Download
Paper Nr: 17
Title:

Sample-based Uncertainty Quantification with a Single Deterministic Neural Network
Authors:

Takuya Kanazawa and Chetan Gupta
Abstract: Development of an accurate, flexible, and numerically efficient uncertainty quantification (UQ) method is one of fundamental challenges in machine learning. Previously, a UQ method called DISCO Nets has been proposed (Bouchacourt et al., 2016) that trains a neural network by minimizing the so-called energy score on training data. This method has shown superior performance on a hand pose estimation task in computer vision, but it remained unclear whether this method works as nicely for regression on tabular data, and how it competes with more recent advanced UQ methods such as NGBoost. In this paper, we propose an improved neural architecture of DISCO Nets that admits a more stable and smooth training. We benchmark this approach on miscellaneous real-world tabular datasets and confirm that it is competitive with or even superior to standard UQ baselines. We also provide a new elementary proof for the validity of using the energy score to learn predictive distributions. Further, we point out that DISCO Nets in its original form ignore epistemic uncertainty and only capture aleatoric uncertainty. We propose a simple fix to this problem.
Download
Paper Nr: 19
Title:

Designing RNA Sequences by Self-play
Authors:

Stephen Obonyo, Nicolas Jouandeau and Dickson Owuor
Abstract: Self-play (SP) is a method in Reinforcement Learning (RL) where an agent learns from the environment by playing against itself until the policy and value functions converge. The SP-based methods have recorded state-of-the-art results in playing different computer games such as Chess, Go and Othello. In this paper, we show how the RNA sequence design problem where a sequence is designed to match a given target structure can be modelled through the SP while performing the state-value evaluation using a deep value network. Our model dubbed RNASP recorded the best and very competitive results on the benchmark RNA design datasets. This work also motivates the application of the self-play to other Computational Biology problems.
Download
Short Papers
Paper Nr: 3
Title:

Signal Detection for Tracer-Based-Sorting using Deep Learning and Synthetic Data
Authors:

Christian Linder, Frank Gaibler, Andreas Margraf and Steffen Geinitz
Abstract: Increasing environmental awareness and new regulations require an improvement of the waste cycle of plastic packaging. Tracer-Based-Sorting (TBS) technology can meet these challenges. Previous studies show the market potential of the technology. This work improves on the solution approach using artificial intelligence to maximize the number of tracers that can be detected accurately. A convolutional neural network and random forest classifier are compared for classification of each tracer based on signal intensities. The approach is validated on different settings using synthetic data to counter the low amount of available data. The results show that theoretically up to 120 tracers can be classified simultaneously under near-optimal conditions. Under more difficult conditions, the maximum number of tracers is reduced to 45. Thus, the approach can increase the diversity of TBS by increasing the maximum tracer count and enable a broader range of applications. This helps to establish the technology in the field of recycling.
Download
Paper Nr: 5
Title:

A Comparative Study of Graph Neural Network Speed Prediction during Periods of Congestion
Authors:

Marko C. Oosthuizen, Alwyn J. Hoffman and Marelie H. Davel
Abstract: Traffic speed prediction using deep learning has been the topic of many studies. In this paper, we analyse the performance of Graph Neural Network-based techniques during periods of traffic congestion. We first compare a selection of recently proposed techniques that claim to achieve good results using the METR-LA and PeMS-BAY data sets. We then investigate the performance of three of these approaches – Graph WaveNet, Spacetime Neural Network (STNN) and Spatio-Temporal Attention Wavenet (STAWnet) – during congested periods, using recurrent congestion patterns to set a threshold for general congestion through the entire traffic network. Our results show that performance deteriorates significantly during congested time periods, which is concerning, as traffic speed prediction is usually of most value during times of congestion. We also found that, while the above approaches perform almost equally in the absence of congestion, there are much bigger differences in performance during periods of congestion.
Download
Paper Nr: 6
Title:

Par-VSOM: Parallel and Stochastic Self-organizing Map Training Algorithm
Authors:

Omar X. Rivera-Morales and Lutz Hamel
Abstract: This work proposes Par-VSOM, a novel parallel version of VSOM, a very efficient implementation of stochastic training for self-organizing maps inspired by ideas from tensor algebra. The new algorithm is implemented using parallel kernels on GPU accelerators. It provides performance increases over the original VSOM algorithm, PyTorch Quicksom parallel version, Tensorflow Xpysom parallel variant, as well as Kohonen’s classic iterative implementation. Here we develop the algorithm in some detail and then demonstrate its performance on several real-world datasets. We also demonstrate that our new algorithm does not sacrifice map quality for speed using the convergence index quality assessment.
Download
Paper Nr: 10
Title:

SENN: Self-evolving Neural Network to Recognize Motor Imagery Thought Patterns
Authors:

Stuti Chug and Vandana Agarwal
Abstract: The EEG-based motor imagery task classification has been a challenge for researchers due to the complex nature of EEG data. Human thoughts are a complex combination of different body limb activations and it is difficult to capture only one thought at a time. The data belonging to different motor imagery thought classes are also not separable linearly. In this paper, a novel technique for efficient and improved motor imagery task classification is proposed. Two major issues in motor imagery task classification of EEG data are addressed - channel selection and radial basis function neural network centers. The channel selection is posed as a combinatorial problem and an evolutionary nature-inspired algorithm PSOCS is proposed to select the most informative and discriminative channels using the Particle Swarm Optimization algorithm. The features are extracted using the selected channels and are subjected to classification. In this paper, a self-evolving radial basis functions neural network (SENN) is proposed based on sub-clusters within each motor imagery task class. The number, centers, and spread of hidden neurons are obtained by the k-means clustering algorithm. The proposed algorithm is validated using the benchmarked datasets BCI Competition IV 2a and BCI Competition IV 2b data set. The proposed technique outperforms some of the existing techniques and classifies the motor imagery tasks efficiently.
Download
Paper Nr: 13
Title:

Suppression of Background Noise in Speech Signals with Artificial Neural Networks, Exemplarily Applied to Keyboard Sounds
Authors:

Leonard Fricke, Jurij Kuzmic and Igor Vatolkin
Abstract: The importance of remote voice communication has greatly increased during the COVID-19 pandemic. With that comes the problem of degraded speech quality because of background noise. While there can be many unwanted background sounds, this work focuses on dynamically suppressing keyboard sounds in speech signals by utilizing artificial neural networks. Based on the Mel spectrograms as inputs, the neural networks are trained to predict how much power of a frequency inside a time window has to be removed to suppress the keyboard sound. For that goal, we have generated audio signals combined from samples of two publicly available datasets with speaker and keyboard noise recordings. Additionally, we compare three network architectures with different parameter settings as well as an open-source tool RNNoise. The results from the experiments described in this paper show that artificial neural networks can be successfully applied to remove complex background noise from speech signals.
Download
Paper Nr: 18
Title:

Improving an Ensemble of Neural Networks via a Novel Multi-class Decomposition Schema
Authors:

Antonio L. Alfeo, Mario A. Cimino and Guido Gagliardi
Abstract: The need for high recognition performance demands increasingly complex machine learning (ML) architec-tures, which might be extremely computationally burdensome to be implemented in real-world. This issue can be addressed by using an ensemble learning model to decompose the multi-class classification problem into many simpler binary classification problems, e.g. each binary classification problem can be handled via a simple multi-layer perceptron (MLP). The so-called one-versus-one (OVO) is a widely used multi-class decomposition schema in which each classifier is trained to distinguish between two classes. However, with an OVO schema each MLP is non-competent to classify instances of classes that have not been used to train it. This results in classification noise that may degrade the performance of the whole ensemble, especially when the number of classes grows. The proposed architecture employs a weighting mechanism to minimize the contribution of the non-competent MLPs and combine their outcomes to effectively solve the multi-class classification problem. In this work, the robustness to the classification noise introduced by non-competent MLPs is measured to assess in what conditions this translates in better classification accuracy. We test the proposed approach with five different benchmark data sets, outperforming both the baseline and one state-of-the-art approach in multi-class decomposition algorithms.
Download
Paper Nr: 20
Title:

Deep Learning of Structural Changes in Historical Buildings: The Case Study of the Pisa Tower
Authors:

Mario A. Cimino, Federico A. Galatolo, Marco Parola, Nicola Perilli and Nunziante Squeglia
Abstract: Structural health monitoring of buildings via agnostic approaches is a research challenge. However, due to the recent advent of pervasive multi-sensor systems, historical data samples are still limited. Consequently, data-driven methods are often unfeasible for long-term assessment. Nevertheless, some famous historical buildings have been subject to monitoring for decades, before the development of smart sensors and Deep Learning (DL). This paper presents a DL approach for the agnostic assessment of structural changes. The proposed approach has been experimented to the stabilizing intervention carried out in 2000-2002 on the leaning tower of Pisa (Italy). The data set is made by operational and environmental measures collected from 1993 to 2006. Both conventional and recent approaches are compared: Multiple Linear regression, LSTM and Tansformer. Experimental results are promising, and clearly shows a better change sensitivity of the LSTM, as well as a better modeling accuracy of the Transformer.
Download
Paper Nr: 23
Title:

Pipeline for Visual Container Inspection Application using Deep Learning
Authors:

Guillem Delgado, Andoni Cortés and Estíbaliz Loyo
Abstract: Containerized cargo transportation systems are associated to many visual inspection tasks. Especially during the process of loading and unloading containers from and to the vessel. More and more of these tasks are being automatized in order to speed up the overall process of transportation. This need for optimized processes calls for new vision systems based on the latest technologies to reduce operation times. In this paper, we propose a pipeline and a complete study of each of its parts in order to provide an end-to-end system that solves and automatizes the process of inspection of a loading or unloading freight container from and to the vessel. We outline all the components involved in a separated way. Tackling from the acquisition of the images at the beginning of the process, to visual inspection tasks such as containers’ id detection, text recognition, damage classification or International Maritime Dangerous Goods (IMDG) detection. In addition, we also propose a heuristic algorithm that is capable of managing all the information from the multiple tasks in order to provide as much insights as possible out of the system.
Download
Paper Nr: 1
Title:

A Haptic Interface for Guiding People with Visual Impairment using Three Dimensional Computer Vision
Authors:

Kevin Swingler and Chris Grigson
Abstract: Computer vision technology has the potential to provide life changing assistance to blind or visually impaired (BVI) people. This paper presents a technique for locating objects in three dimensions and guiding a person’s hand to the object. Computer vision algorithms are used to locate both objects of interest and the user’s hand. Their relative locations are used to calculate the movement required to take the hand closer to the object. The required direction is signaled to the user via a haptic wrist band, which consists of four haptic motors worn at the four compass points on the wrist. Guidance works both in two and three dimensions, making use of both colour and depth map inputs from a camera. User testing found that people were able to follow the haptic instructions and move their hand to locations on vertical or horizontal surfaces. This work is part of the Artificial Intelligence Sight Loss Assistant (AISLA) project.
Download
Paper Nr: 11
Title:

Investigating Prediction Models for Vehicle Demand in a Service Industry
Authors:

Ahmed Alzaidi, Siddhartha Shakya and Himadri Khargharia
Abstract: Demand prediction is an important part of resource management. Higher forecasting accuracy leads to better decision taking capabilities, especially in a competitive service-based business such as telecommunication services. In this paper, a telecommunication service provider’s data on the use of vehicles by their employees is analyzed and used to forecast the vehicle booking demand for the future at different geographical locations. We implement multiple forecasting models and investigate the effect on forecasting accuracy of two prediction strategies, namely the Direct multi-step forecasting strategy (DMS) and the Rolling mechanism strategy (RMS). Moreover, the effect of different external inputs such as temperatures and holidays were tested. The results show that both DMS and RMS can be used to forecast vehicle demand, with the highest improvement in forecasting achieved through the addition of the holiday input, particularly by using the RMS strategy in the majority of the cases.
Download
Paper Nr: 14
Title:

Technology Transfer of Convolutional Neural Networks: An Example
Authors:

Thanakij Wanavit, Samuel Sallee, Chedtha Puncreobutr, Leslie Klieb and Pin P. Tea-Makorn
Abstract: A number of university groups have shown that neural networks, especially U-nets, can satisfactorily segment CT-scans of bones. Segmentation, labelling the scans where bone and enamel are and where not, can be used to make a 3D model of the skull. This paper gives an overview of efforts to transfer university-based research work for use to a company that manufactures titanium meshes for brain surgery. It discusses issues and pitfalls in such a transition. A working prototype is discussed.
Download
Paper Nr: 15
Title:

A MLP for Dryer Energy Consumption Prediction in Wood Panel Industry
Authors:

Valentin Chazelle, Philippe Thomas, Hind B. El-Haouzi and Christophe Heleu
Abstract: The drying operation is the most energy consuming step of particle board manufacturing process. Even if a great academic and industrial effort has been furnished for last years, the prediction of this energy consumption is still a challenging issue. This paper deals with the energy consumption prediction for industrial wood drying. The study of an European particle board manufacturer’s industrial dryers has provided data sets for two both fresh and recycled wood drying processes. Based on these, MLP Neural network models have been developed and tested. Several tests have been conduced to identify and select the best MLP model’s structure to find a satisfying trade-off between model accuracy and maintenance efficiency. The proposed MLP models have either been distinctly trained on the datasets from both the first and second dryers, and then on their combination, in order to increase data diversity and to reduce training time and model maintenance. Then, the neural network based on the merged dataset has been compared to those developed from the single datasets. This experiment led to the conclusion that, the construction of a global model representing the operation of the two dryers is less accurate than the construction of a dedicated model for each dryer. Yet, the performances of combination model remain acceptable.
Download