Browsar Examensarbeten för masterexamen // Master Theses efter Program "Data science and AI (MPDSC), MSc"
Visar 1 - 15 av 15
Sökresultat per sida
- PostAn Empirical Survey of Bandits in an Industrial Recommender System Setting(2023) Brandby, Johan; Schwarz, Tobias; Chalmers tekniska högskola / Institutionen för data och informationsteknik; Chalmers University of Technology / Department of Computer Science and Engineering; Dubhashi, Devdatt; Jorge, EmilioIn this thesis, the effects of incorporating unstructured data—images in the wild—in contextual multi-armed bandits are investigated, when used within a recommender system setting, which focuses on picture-based content suggestion. The idea is to employ image features, extracted by a pre-trained convolutional neural network, and study the resulting bandit behaviors when including respective excluding this information in the typical context creation, which normally relies on structured data sources—such as metadata. The evaluation is made both online, through A/B-testing enabled by the industrial partner YouPic AB, and offline, effectuated by a simulation pipeline that models the online counterpart. The results are compiled as a survey, covering a selection of contextual bandit algorithms, highlighting the differences brought by the unstructured data. The offline result points towards that if the contextual bandit utilizes a joint or hybrid action-value function, with respect to the parameterization, the addition of the image vectors can significantly outperform the instances without it; however, if a disjoint model is instead employed, no noticeable change is observed. In comparison, those from the online trials can be interpreted as supporting the inclusion of convolutional features, but due to meager and unbalanced sample sizes, the outcomes are deemed inconclusive. To summarize, though there is support for incorporating unstructured data, given that the action-value function is joint or hybrid, the online experiments gave too little evidence for any trustworthy findings; in other words, the question is still partially open.
- PostApplicability of Supervised Machine Learning for CI Configuration Selection(2023) Lönnfält, Albin; Tu, Viktor; Chalmers tekniska högskola / Institutionen för data och informationsteknik; Chalmers University of Technology / Department of Computer Science and Engineering; Gren, Lucas; Gay, GregoryThis study introduces a novel supervised machine learning (ML) model for accurately assigning CI configurations to test specifications. Current solutions to optimize selection of CI configurations lack the ability to select CI configurations for individual test cases and assigning them into predefined CI configurations. The model employs an ensemble architecture with three sub-models and a rule-based component, each focusing on specific aspects of the problem. Extensive model analysis reveals important features that contribute to the assignment process. A decision support system based on the ML model is developed to evaluate the applicability of supervised ML in CI configuration assignment, validated through a survey study involving domain experts. The study demonstrates that supervised ML can exceed the performance requirements of domain experts. Certain features in test specifications are found to be influential in the assignment outcome. Implementing supervised ML brings business value, reducing misassignments, saving time, and reducing fault slip through. Proposed future research includes exploring fully automated CI configuration assignments and investigating more complex ML models, such as neural networks, for enhanced performance and exploring the potential for fully automated adaptation.
- PostAutoencoder and Active Learning to Reduce False Positive Warnings in a Slippery Road Alert System(2023) Axenhamn, Philip; Greppe, Andreas; Chalmers tekniska högskola / Institutionen för data och informationsteknik; Chalmers University of Technology / Department of Computer Science and Engineering; Damaschke, Peter; Selpi, SelpiModern vehicles are commonly equipped with a slippery road alert (SRA) system that warns drivers of slippery roads. Current implementations of this system occasionally produce warnings when the road is not slippery. These warnings are called false positives and can harm the system’s trustworthiness. In this thesis, we propose a false positive filter capable of reducing the number of false positive alerts generated by an SRA system based on two machine learning techniques: autoencoder and active learning. The available data was completely unlabeled and contained few informative features. The analysis of this data showed that vehicles send bursts of data points forming sequences. Due to the limitations of the data, multi-variate time series were constructed with the idea that the sequential data might reveal more about the situation than a single-point measurement. Furthermore, the sequences were grouped into true- and false-positive classes based on assumptions of the causes for the alerts, such as driving on ice or a speed bump. The sequential data was used to train GRU- and LSTM-based autoencoders and classifiers to detect sequences that correspond to false positive situations such that they can be removed. The hyperparameters for the models were tuned using Optuna and the best-performing models with the most optimal hyperparameters were further evaluated. Since the data was not labeled, the actual performance of the proposed solution could not be assessed. Instead, the evaluation was based on computing the proportion of remaining assumed true positive (ATP) sequences and assumed false positive (AFP) sequences after filtering. The results show that the LSTM autoencoder could find patterns in the sequential data and was able to remove 43% of the AFP sequences while retaining 90% of the ATP sequences. The active learning approach proved to not work well with the available data.
- PostCombinatorial Optimization with Reinforcement Learning(2023) Persson Hijazi, Aladdin; Persson, Sanna; Chalmers tekniska högskola / Institutionen för data och informationsteknik; Chalmers University of Technology / Department of Computer Science and Engineering; Bernardy, Jean-Philippe; Damaschke, PeterThis master’s thesis delves into the topic of solving combinatorial optimization problems with methods based on reinforcement learning, and specifically, we explore the potential of iterative route decoding and gradient updates in enhancing the performance of route decoding. In this context, route decoding refers to determining the most efficient route for a set of destinations, a combinatorial optimization problem often encountered in logistics and transportation planning. We introduce two methods for iteratively updating solutions for the heterogeneous capacitated vehicle routing problems. They are built upon a reinforcement learning algorithm with an attention graph encoder and use previously computed routes for an instance to improve solution quality. Our results show improved performance, in particular, on out-of-distribution data, which suggests the practical applicability of the methods. In particular, our results show that a pre-trained route planner can, with a few gradient updates with a policy gradient method, significantly improve on out-ofdistribution data.
- PostCut-in Detection from Videos(2023) Udayakumar, Apoorva; Varma, Aditya Padmanabhan; Chalmers tekniska högskola / Institutionen för data och informationsteknik; Chalmers University of Technology / Department of Computer Science and Engineering; Damaschke, Peter; Panahi, AshkanFor autonomous driving, it is crucial to anticipate the behaviour of other road users and act accordingly. One important scenario is when a vehicle cuts into the lane of the ego-vehicle with or without sufficient indicator cue. This thesis studies such a scenario and investigates the application of deep learning techniques for understanding and predicting when the cut-in maneuver is performed by the vehicles ahead. As a first step, indicator cues from videos of vehicles performing cut-ins are detected successfully with an F1 score of of 83% and recall value of 85%. We achieve this result by employing ResNet-18 and CNN-LSTM with a tuned level of context around the target vehicle. Further we predict the estimates of interest such as start and end of cut-in intention using the same architectures and discuss the challenges.
- PostData driven running technique identification(2023) Lamm, Johan; Preiman, Jessica; Chalmers tekniska högskola / Institutionen för data och informationsteknik; Chalmers University of Technology / Department of Computer Science and Engineering; Dubhashi, Devdatt; Johansson, MoaWe evaluate if acceleration and rotational IMU data and marker-based positional data can be used to quantify the running technique. We also investigate patterns between the runners’ anatomy, fitness level and technique, and how common instructions impact their running technique. Running technique data, consisting of rotational velocity and acceleration from a foot mounted IMU and position from a marker-based motion capture system, is collected from 47 participants, together with data on anatomy and fitness level. Participants perform a testing protocol containing treadmill running at different velocities while receiving different technique instructions. Data is processed to extract a representative stride cycle for each data source for every participant at every velocity and technique instruction. We evaluate three methods to quantify the technique using dimensionality reduction and reconstruction: sequential feature selection using multivaritate linear regression, principal component analysis, and autoencoder. Best performance is obtained for principal component analysis on all data sources. Information loss is significantly larger on rotational and acceleration data from the IMU than for positional data from the marker-based system. Limited patterns between the anatomy and fitness level of runners and their technique were observed, and the found patterns are generally on parameters that are not related to technique, such as ground contact time and contact to flight-time ratio. Most technique instructions are shown to impact technique, but the effect diminishes as velocity increases. A larger impact is seen when runners are asked to increase back-kick height, knee lift, or frequency, and a smaller impact is seen when asked to land further back with the foot or push the hip forward.
- PostDeep Dynamic Graphical Models for Molecular Kinetics(2023) Gao, Wenli; Su, Enmin; Chalmers tekniska högskola / Institutionen för data och informationsteknik; Chalmers University of Technology / Department of Computer Science and Engineering; Damaschke, Peter; Olsson, SimonWith the massive growth of molecular dynamics simulation results comes a great demand for efficient analysis methods to distill essential information from simulation and enable quantitative characterization of molecular properties. Dynamic Graphical Model (DGM) is currently the most data-efficient method towards this goal. However, DGMs rely on extensive manual intervention by experts: division of molecules into smaller subsystems and their discretization into an unknown number of states. We aim to automate this expert-guided procedure using a deep-learning approach and make an end-to-end learning system. To achieve this, we examine the Variational Approach to Markov Processes (VAMP), and its ability to detect meta-stable subsystems in molecular systems, and to decide the number of states for each subsystem. We put forward a model which uses VAMP to learn subsystem states via a deep neural network and DGM to connect the subsystems by modeling their time-correlated dynamics. The model is trained in an end-to-end manner and optimized using a weighted sum of VAMP loss, DGM loss, and regularizer. We also introduce a pruning-based algorithm to automatically decide the number of states per subsystem. Our results show that VAMP is suitable for enumerating subsystems of a molecular system, however, VAMP alone cannot decide the number of states for each subsystem. This thesis sheds light on how an end-to-end learning system may be built with DGMs to analyze molecular dynamics and outlines possible future extensions of this work.
- PostDetecting Metastable States in Proteins using E(3) Equivariant VAMPnets(2023) Arnesen , Sara; Nordström, David; Chalmers tekniska högskola / Institutionen för data och informationsteknik; Chalmers University of Technology / Department of Computer Science and Engineering; Dubhashi, Devdatt; Olsson, SimonAs proteins fold, they encounter intermediary conformations, often denoted metastable states, that are vital to deciphering diseases related to malfunctions in conformational changes. To detect these metastable states, a deep learning framework using the variational approach for Markov processes (VAMP) has been proposed, dubbed VAMPnets. In this master’s thesis, we improve the training of VAMPnets through the use of E(3) equivariant neural networks. These networks incorporate the symmetries of Euclidean space, facilitating faster and more data-efficient learning. To study the effectiveness of these networks, we benchmark two different equivariant Transformer architectures and an equivariant convolutional network against both a simple and an invariant multilayered perceptron. The models are evaluated on molecular dynamics trajectories of alanine dipeptide and protein folding datasets. The use of E(3) equivariant neural networks in training VAMPnets is shown to significantly improve the prediction accuracy on random downsampled data. Using only 1% of the dataset, the equivariant Transformer achieves almost twice the VAMP-2 score as the benchmarks. Furthermore, the model exhibits improved robustness. With only 20% data remaining, the model scores on par with the complete dataset. On average, the model requires significantly fewer backward passes, converging more than twice as fast as the benchmark models, showing enhanced data efficiency. Furthermore, the results highlight the significant computational burden that equivariant neural networks pose, especially for larger molecules, proving almost 1,000 times slower on the protein folding dataset. Finally, we propose a novel algorithm for detecting the number of metastable states of a molecule using the VAMP-2 score and provide estimates for the 12 proteins in the protein folding dataset.
- PostImportance Sampling in Deep Learning Object Detection - An empirical investigation into accelerating deep learning object detection training by exploiting informative data points(2023) Huang, Alexander; Kvernes, Johannes; Chalmers tekniska högskola / Institutionen för data och informationsteknik; Chalmers University of Technology / Department of Computer Science and Engineering; Angelov, Krasimir; Yu, YinanIn recent years, the field of deep learning object detection has witnessed a notable surge in progress, largely fueled by the growth of available data. While the ever growing amount of data has enabled complex object detection models to improve generalization performance and to facilitate training phases that are less prone to over-fitting, the time to reach desired performance with this remarkable amount of data has been reported to be an emerging problem in practice. This is certainly the case when each sample consist of high resolution image data that could burden the capacity of loading data. This thesis explores the possibilities of leveraging importance sampling as a means to accelerate gradient-based training of an object detection model. This avenue of research aims at biasing the sampling of training examples during training, in the hope of exploiting informative samples and reduce the amount of computation on uninformative, noisy and out-of-distribution samples. While previous art shows compelling evidence of importance sampling in deep learning, it is consistently reported in a context of less complex tasks. In this work, we propose techniques that can be applied to single-stage anchor-free object detectors, in order to adopt importance sampling to accelerate training. Our methods do not require modifications to the objective function, and allow for a biased sampling procedure that remains consistent across runs and samples. Our results suggest that an uncertainty-based heuristic outperforms loss-based heuristics, and that object detection training can be subject to a remarkable speed-up in terms of reaching the baseline’s performance in fewer iterations, where the baseline samples the training examples uniformly without replacement. Furthermore, the empiric observations reported in this work also indicate that an increased final generalization performance can be achieved given an equal amount of training time when compared to the baseline.
- PostLow latency video analytics system with multi-exit neural networks(2022) HARINDRAN, NEETHU; POOJARY, BHARATH; Chalmers tekniska högskola / Institutionen för data och informationsteknik; Chalmers University of Technology / Department of Computer Science and Engineering; Tsigas, Philippas; Ali-Eldin Hassan, AhmedComputer vision-based control systems have become increasingly powerful and promising in tackling real-world problems. This can be accredited to the use of deep learning methods in these systems with state-of-the-art performance sometimes outperforming humans in tasks which require subjective decision making. This has resulted in increased interest in these systems from Swedish industry, including Volvo. One example system where these systems are used is the Volvo GPSS system, where semantic segmentation is used to perform real-time decisions based on pixel level classification of a monitored area. However, such systems frequently deal with a trade-off between latency and accuracy. This is primarily due to the increasing number of model layers being used to develop Deep-Neural-Network models for vision systems, resulting in equal resource utilization regardless of input complexity. In this thesis, we develop an approach that employs input adaptive multi-exit strategy to exploit latency benefits of dynamic processing based on the input complexity. The proposed approach aims to have a reduced average inference time as the simple input samples takes an early exit and only the complex samples need more computation offered by all the model layers. The open source CityScapes dataset and the Volvo dataset were used in a number of multi-exit semantic segmentation experiments with HRNet architecture chosen as the backbone. The thesis work studies three novel exit strategies, including reinforcement learning, auxiliary models, and fast Fourier transform. Out of all the methods examined, the reinforcement learningbased exit strategy displayed the best performance advantages, with accuracy on par with unbranched HRNet and a significant decrease in latency and computation.
- PostMachine Learning for Detecting Gender Bias at Chalmers(2023) Nilsson, Linnea; Lindau, Sarah; Chalmers tekniska högskola / Institutionen för data och informationsteknik; Chalmers University of Technology / Department of Computer Science and Engineering; Johansson, Moa; Ljunglöf, PeterThis thesis studies gender bias in course evaluations through the lens of machine learning and NLP. Different methods are used to examine and explore the data and find differences in what students write about courses depending on the gender of the examiner. The data is also examined using more traditional statistical methods to get an understanding of how the students’ impressions of the courses are related to the gender of the examiner. Other aspects related to gender and gender bias are also examined, such as how the proportion of female students relates to the gender of the examiner and whether male or female examiners give different grades to their students. Student grades and teaching language are also factors that are being examined to see whether there is any bias against female examiners or students that is easily detectable in the data. The main findings are that courses with female examiners seem to get lower overall impression scores than those with male examiners. Courses taught in Swedish also receive lower scores, compared to the English courses. No clear patterns as to what words are used when writing comments about a course with a male or female examiner were found. When trying to predict the author gender the patterns were clearer, finding that men write more words directly related to the course and women write more words related to communication.
- PostMachine Learning for generative painting informed by visual arts(2023) Wang, Chaoming; Chalmers tekniska högskola / Institutionen för data och informationsteknik; Chalmers University of Technology / Department of Computer Science and Engineering; Dahlstedt, Palle; Tatar, KıvançVisual art practice is a complicated, varied, creative process based on the artist’s style and preferences. Although many studies have attempted to apply artificial intelligence techniques to art production and statistical analysis, there is still significant scope for exploring how to incorporate the techniques in visual arts practices into generative painting pipelines using Machine Learning. This thesis applies machine learning to analyzing painting techniques in painting practices with a research-through-design approach. The problem is mainly presented as tasks such as segmentation of artworks (in this thesis, paintings), stroke prediction, and the presentation of painting processes based on different painting techniques through different algorithmic pipelines. The results show that most segmentation models based on photo training are challenging to apply to the segmentation of artwork components directly, and relevant improvement solutions are discussed in Chapter 6. In addition, due to the diverse presentation of painting art, this paper presents different painting techniques based on the foreground and background segmentation and ’blocking-in’ techniques based on line detection. It discusses the possibility of transferring these painting processes to other painting processes.
- PostMachine Learning for Predicting Targeted Protein Degradation(2023) Ribes, Stefano; Chalmers tekniska högskola / Institutionen för data och informationsteknik; Chalmers University of Technology / Department of Computer Science and Engineering; Damaschke, Peter; Mercado, RocíoPROteolysis TArgeting Chimeras (PROTACs) are an emerging high-potential therapeutic technology. PROTACs leverage the ubiquitination and proteasome processes within a cell to degrade a Protein Of Interest (POI). Designing new PROTAC molecules, however, is a challenging task, as assessing the degradation efficacy of PROTACs often requires extensive effort, mostly in terms of expertise, cost and time, for instance via laboratory assays. Machine Learning (ML) and Deep Learning (DL) technologies are revolutionizing many scientific fields, including the drug development pipeline. In this thesis, we present the data collection and curation strategy, as well as several candidate DL models, for ultimately predicting the degradation efficacy of PROTAC molecules. In order to train and evaluate our system, we propose a curated version of open source datasets from literature. Relevant features such as pDC50, Dmax, E3 ligase type, POI amino acid sequence, and experimental cell type are carefully organized and parsed via a Named Entity Recognition system based on a BERT model. The curated datasets have been used for developing three candidate DL models. Each DL model is designed to leverage different PROTAC representations: molecular fingerprints, molecular graphs and tokenized SMILES. The proposed models are evaluated against an XGBoost model baseline and the State-of-The-Art (SOTA) model for predicting PROTACs degradation activity. Overall, our best DL models achieved a validation accuracy of 80.26% versus SOTA’s 77.95% score, and a Receiver Operating Characteristic Area Under the Curve (ROC AUC) validation score of 0.849 versus SOTA’ 0.847.
- PostModel-based deadlock prevention for traffic planning of autonomous vehicles(2023) Möller, David; Ohlin, Alexander; Chalmers tekniska högskola / Institutionen för data och informationsteknik; Chalmers University of Technology / Department of Computer Science and Engineering; Damaschke, Peter; Gheorghiu, AndruVolvo Autonomous Solutions are developing a system for planning the routes of fleets of autonomous vehicles. Autonomous control creates several problems that must be solved; among these is the possibility for the policy of said vehicles to end up in deadlock. This thesis proposes new concepts to describe the problem and methods for preventing a vehicle fleet from deadlocking. As the action that led to deadlock might not be recent, the term implicit deadlock was introduced, which is a configuration of vehicle positions from which deadlock is inevitable. The methods developed successfully prevent deadlocks at several pilot and test sites. However, results indicate that time for computing implicit deadlocks grows exponentially in the size of the site and the number of vehicles in the fleet. A neural network model was also trained using data generated from preprocessing of deadlocks to approximate the process and enable deadlock predictions not discovered before.
- PostPredictive Maintenance in Production Robots in a Real World Industrial Setting(2023) Wennerström, Karl; Svensson, Adam; Chalmers tekniska högskola / Institutionen för data och informationsteknik; Chalmers University of Technology / Department of Computer Science and Engineering; Sheeran, Mary; Yu, YinanWith the exponential growth of data and advancements in AI technology, Predictive Maintenance (PdM) has emerged as a vital practice for optimizing equipment maintenance and minimizing unplanned downtime. This study was performed in collaboration with Sandvik Coromant, a company producing steel products and actively collecting data for evaluation purposes, to investigate how the collected data can be utilized for decision-making processes. Specifically, the study analyses vibrational data to address the issues related to unexpected tool malfunctions. Based on the available data, anomaly detection was identified as the most suitable approach to leverage the stored data based on its characteristics. A comparative study where various anomaly detection models were evaluated demonstrated that a reconstruction-based LSTM autoencoder yields the highest performance. The reconstruction approach exhibited its effectiveness in detecting and flagging potential abnormalities, capturing 71% of the malfunctions with an F1 score of 0.75 for the data used for the comparison. Extending the model to other tools displayed the challenges posed within time-series analysis, proving unique characteristics for each case. The findings from this study provide valuable insights into the implementation of anomaly detection techniques for leveraging collected data and enhancing decision-making processes in Sandvik Coromant and similar industrial settings.