Welcome to the Radboud Educational Repository
Here the Radboud University presents theses written by students affiliated with the various bachelor and master programmes offered at the Radboud University, as well as papers written by students of the Radboud Honours Academy.
ItemDiscovering new intermediate variables in Bayesian networks(2021-06-01)Bayesian networks have been used to model various types of cognitive process- ing, but have so far lacked structure learning capabilities, in particular the capa- bility to discover and integrate new variables. The paper proposes a method that automatically adds new intermediate variables that compress information coming into nodes, based on the properties of probability distributions in the network. This serves as a proof-of-concept method of adding new variables to Bayesian net- works. To add nodes in an informed manner, the method utilizes metrics from information theory such as entropy and Kullback-Leibler divergence. ItemSelf-Attention in Convolutional Neural Networks: a Potentially Fundamental Improvement for Image Classi cation(2021-06-18)Conventional (feedforward) Deep Neural Networks (DNNs) have some in- herent limitations and weaknesses that humans do not have. In the domain of computer vision, this becomes clear when a CNN is confronted with ad- versarial attacks. An explanation for this di erence in performance and robustness between a CNN and the human brain could be the usage of recurrence. The visual cortex of the human brain amply uses feedback con- nections when processing visual input. Most conventional CNNs, however, do not make use of this type of connections. Therefore, we believe that recurrence might help improve performance of CNNs, at least in certain challenging tasks. In this paper, we propose a kind of recurrence that is based on top-down attention. For our experiments, we used a feedforward CNN as a benchmark network and compared it to an equivalent network but with top-down attention. We compared the performance of both models in a digit classi cation task that was complicated by partial occlusion. In this task, the recurrent network signi cantly outperformed the benchmark net- work. Although further research is needed to make de nitive conclusions, the results provide promising initial evidence that the use of recurrence in a CNN can lead to fundamental improvements in image classi cation. ItemExamining Human Walking Characteristics Using Video-Based Motion Tracking(2021-06-18)Markerless gait recognition is a fast-evolving eld. In the study of human movement, it can be a great asset and be preferred over marker-based meth- ods, because markers themselves are obtrusive and may be inaccurate. The fast development in this eld is marked by innovations such as DeepLab- Cut, a markerless pose estimation method based on transfer learning with deep neural networks that approaches human-level labeling accuracy with minimal training data. In order to test the viability of DeepLabCut, several gait parameters were identi ed from videos of walking participants to repro- duce known di erences in these parameters between men and women. These parameters are walking velocity, step frequency, and step length, in which previous research has shown signi cant di erences between men and women, namely that men walk with greater velocity, with greater step length, but with a lower frequency. Using a well-known dataset consisting of people walking perpendicular to the camera, it was found that men walk with greater velocity and greater step length, but it was not found that women walk with greater frequency. These results have various consequences to the implementations of gait-identifying software, and the development of systems such as DeepLabCut. In the improvement of our understanding of human walking, this research could be expanded to investigate the effects of these parameters on the energy consumption of walking participants. ItemUsability of CNN and Attention Mechanisms for Classifying Melanoma Image(2021-07-02)Malignant melanoma accounts for about 2% of all malignancies in the West- ern countries, particularly in the United States, and is a disease that kills more than 9,000 people each year. In general, skin lesions are di cult to detect accurately through visual criteria, but if they are detected well at an early stage, unnecessary time and cost for additional diagnosis can be reduced. This study proposes a solution using a deep learning-based CNN to solve the problem of skin cancer classi cation. Preprocessing to solve the class imbalance problem is performed and transfer learning architecture to select a backbone architecture model and train it successfully. Furthermore, we apply one of the new deep learning techniques, so-called 'Attention', to the existing model to nd out whether the model architecture replaced by the attention layer has better performance. As a result, it is expected that several proposed arti cial intelligence algorithms will be utilized to build better computer-aided diagnostic algorithms, which will help early detec- tion of malignant melanoma. ItemThe role of working memory mechanisms in speech recognition(2021-07-02)Spiking neural networks are build from relevant neuro-biologically inspired mechanisms. Complex dynamics and network structures create models that can be used to make inferences about the existing knowledge pool on psycho-linguistics and neuroscience. This paper discusses working memory mechanisms and how they can be implemented in spiking neural models for a speech recognition task. Several mechanisms are discussed, that are all based on known phenomena in the brain. By comparing classi cation scores on states of the network during word representation, the performance of models with several working memory mechanisms is compared. These mechanisms are neuronal adaptation, implemented with an adaptive current, and short-term synaptic plasticity, implemented through a phenomenological model. It is concluded that the dynamics of the neuronal adaptation and short-term synaptic plasticity interact within the network. In particular, it is shown that the interactions between these mechanisms have di ering e ects on learning in the network.