Abstract This paper presents a novel sampling scheme for masked non-autoregressive generative modeling. We identify the limitations of TimeVQVAE, MaskGIT, and Token-Critic in their sampling processes, and propose Enhanced Sampling Scheme (ESS) to overcome these limitations. ESS explicitly ensures both sample diversity and fidelity, and consists of three stages: Naive Iterative Decoding, Critical Reverse Sampling, and Critical Resampling. ESS starts by sampling a token set using the naive iterative decoding as proposed in MaskGIT, ensuring sample diversity.
Abstract Modelling the behaviour and performance of cyclists during competition ahead of time has several useful applications to professional cycling teams. Although existing physics-based approaches perform reasonably well in this regard, they generally require several assumed parameters. Performance is also dependant on several variables which may be hard to quantify, such as a rider’s cornering ability and pacing strategy. The work presented in this thesis is a fully data-driven approach to speed pre- diction in time trials performed by elite cyclists.
Abstract In the telecommunications domain, as network infrastructure grows more complex, it is becoming increasingly important to develop systems that are able to automate efficient and accurate anomaly detection. Such systems operate on Key Performance Indicators (KPIs) collected from network base stations, and are vital to alert for possible critical operational incidents in a timely manner. These KPIs are most commonly in the form of multivariate time-series, which are challenging to perform anomaly detection on, as anomalous patterns occurs both as correlations between features, as well as temporal windows.
Abstract In the following pages lies our master’s thesis on how the recent advances in deep learning architectures, namely graph neural networks, can perform unsupervised anomaly detection in the Telecommunications (telco) domain. This work is moti- vated by the need for efficient and accurate anomaly detection in the telco domain, where Key Performance Indicators (KPIs) of base stations are continuously being monitored. Furthermore, network infrastructures are constantly being upgraded, 5G is on its way, and there is an exponential increase of devices and antennas.
Abstract Self-supervised learning for image representations has recently had many breakthroughs with respect to linear evaluation and fine-tuning evaluation. These approaches rely on both cleverly crafted loss functions and training setups to avoid the feature collapse problem. In this paper, we improve on the recently proposed VICReg paper, which introduced a loss function that does not rely on specialized training loops to converge to useful representations. Our method improves on a covariance term proposed in VICReg, and in addition we augment the head of the architecture by an IterNorm layer that greatly accelerates convergence of the model.
Abstract The aim of this thesis was to predict the wind power production volume of a large geographical region given the Numerical Weather Prediction data (NWP) over the region using deep learning. Accurate production volume predictions is important for power grid balancing, production planning, and price estimation. Having an accurate forecast for the upcoming wind power production volume has become more and more important in the past years due to the fast increasing number of installed wind turbines and installed total production capacity.
Abstract With the growing inclusion of renewable energy sources, developing price models for intraday trading has become an essential task for many market participants in order to optimize the decision-making process. Yet the available literature on the topic has not been keeping up with the pace of increased intraday trading activity. We predict prices in the final hour prior to delivery on the German intraday market, utilizing Deep Learning techniques. This thesis looks into the usage of feed-forward neural networks and recurrent neural networks (LSTM and GRU).
