Neural networks are susceptible to minor distortions in their input, which can lead to errors they would not otherwise make. This susceptibility, termed as the network’s robustness, is a crucial...Show moreNeural networks are susceptible to minor distortions in their input, which can lead to errors they would not otherwise make. This susceptibility, termed as the network’s robustness, is a crucial aspect to evaluate. While several methods exist for measuring robustness, they usually suffer from interpretability issues and do not provide a statistical guarantee. In this work, we propose a novel robustness measure that addresses these short- comings by modeling the robustness as a probability distribution and mea- suring its 0.05 quantile. Additionally, previous work suggests the poten- tial modeling of robustness through a log-normal distribution. To eval- uate this hypothesis and its computational benefits, we introduce an es- timator that assumes the distribution is log-normal. A comparison with the standard parameter-free estimator reveals significantly improved com- putational efficiency with the parametrized approach. However, the log- normal assumption requires further research. The assumption is too strong and needs to be relaxed before the parametrized estimator can reliably be utilized.Show less
The ability to correctly classify an individual’s clinical status could help pave the way for early treatment programs for disorders and illnesses of mental- and physical nature alike. Neuroimaging...Show moreThe ability to correctly classify an individual’s clinical status could help pave the way for early treatment programs for disorders and illnesses of mental- and physical nature alike. Neuroimaging data could serve as a basis in reaching this goal of accurate classification. The use of said data, however, does come with challenges. A prominent one of which is the fact that neuroimaging data is highly dimensional, meaning that the amount of features largely exceeds the number of subjects within the data set. Furthermore, research has indicated that the use of heterogeneous, but complimentary, data derived from multiple modalities can be an asset to a model used in a classification setting (Zhang et al., 2010; Schouten et al., 2016). The challenge arises in how to combine the different modalities within a single model. A solution could be the use of machine learning algorithms which search for patterns in the data to draw conclusions. Current literature is, however, lacking meaningful comparisons between different machine learning techniques. Within this project, three different algorithms (support vector machines, Gaussian process classification and multiple kernel learning) have been selected to get insight into (1) whether or not machine learning is able to cope with challenges in the use of neuroimaging data, (2) the difference in performance between these methods and (3) whether or not the use of multiple modalities leads to better results in classification. To this end, 16 alcohol-dependent respondents have been selected along with 32 age-matched healthy controls and have been subjected to both MRI and PET. Models have been trained on data from both separate modalities and on data combining the two modalities. The performances of the models have been assessed by leave-one-subject-out cross-validation and expressed in balanced accuracy and area under the curve. Results indicate that the chosen methods are effective in overcoming challenges arising in the use of neuroimaging data as a means of classification. High balanced accuracies have been found ranging from 76.56% (GPC using PET data) to 100% (GPC using MRI data). Different situations are cause for different solutions and the right choice of algorithm seems to be dependent on, for instance, the fact if either unimodal- or multimodal data is used. Also, settings/optimization of parameters within the model can make a large impact on accuracy. It is therefore advised that researchers try different algorithms and settings before selecting a technique. Different options need to be weighed in order to receive the best possible outcome.Show less
Electroencephalography (EEG) has been used for over a century to measure electric potentials in the brain. The brain signals can be used in a Brain Computer Interface (BCI) to control machines,...Show moreElectroencephalography (EEG) has been used for over a century to measure electric potentials in the brain. The brain signals can be used in a Brain Computer Interface (BCI) to control machines, like prostheses or exoskeletons. In the last decades, consumer-grade EEG devices have become available. Their performance and reliability, however, does not match that of medical EEG equipment. In this project, the possibilities for classification with such a consumer-grade EEG device, the EMOTIV EPOC+, were explored. The main challenge was to overcome the large noise and nuisances in the EEG signal to be able to classify brain states induced by motoric tasks and sensory stimuli. The effect of several preprocessing techniques was studied: Independent Component Analysis, a denoising autoencoder, and adaptive filtering with an autoencoder. For the classification, various machine learning approaches were proposed to analyse the EEG signals, ranging from a naive Bayes classifier to Deep Convolutional Neural Networks. Furthermore, the effect of motivating the structure of a neural network for this classification task was investigated. It was found that using the spectral features of the EEG signal as a motivation for the structure of neural networks helps classifiers identify brain states that are not classified by other types of neural networks.Show less
