Gideon Jägenstedt defended his Master Thesis on 8 June 2023 at 11:00. Congrats!
Title: End-to-End Object Tracking on Simulated Microscopy Video
Object tracking in microscopy time-lapse videos is currently mostly done in two steps often using two neural network models, first the images are segmented in order to detect each object within each time frame and extract their centroids using one neural network model. A selected set of properties on the centroids are then used as an input to a second neural network that creates the temporal trajectories by linking the centroids over a sequence of frames.
This work proposes a novel method to combine the object detection step and the
linking step which should, in theory, create better linking in time since the combined model has access to not only a set of properties but the complete image of the centroids. Two different architectures of a combined model were tested, one supervised model based on graph neural networks (GNN) and one unsupervised model based on a variational autoencoder (VAE).
The supervised GNN-based model did not succeed in predicting the position of the centroids, but it showed promise in linking the centroids between frames. Therefore, the VAE-based model was developed that uses the same approach for linking. The VAE-based model resulted in a mean absolute error of under 0.002 on its detection placement, a detection miss-rate of 2.69 %, and an F1-score of 81.2 % when linking trajectories on simulated data.
Supervisor: Jesus Pineda Castro Examiner: Giovanni Volpe Opponent: Mirja Granfors
Mirja Granfors defended her Master thesis in physics at the University of Gothenburg on June 8 2023. Congrats!
Title: Enhancing Graph Analysis and Compression with Multihead Attention and Graph-Pooling Autoencoders
Graphs are used to model complex relationships in various domains. Analyzing and classifying graphs efficiently poses significant challenges due to their inherent structural complexity. This thesis presents two distinct projects aimed at enhancing graph analysis and compression through novel and innovative techniques. In the first project, a multihead attention module for node features is developed, enabling effective prediction of graph edges for connection in time. By applying attention mechanisms, the module selectively focuses on relevant features, facilitating accurate edge predictions. This approach expands the potential applications of graph analysis by improving the understanding of graph connectivity and identifying critical relationships between nodes. The second project introduces a novel graph autoencoder with multiple steps of size reduction by graph-pooling. Unlike traditional graph autoencoders, which commonly employ graph convolutional networks, this approach utilizes several poolings to capture diverse structural information and compress the graph representation. The pooling-based autoencoder not only achieves efficient graph compression but also captures the structural information of the graph. This enables the classification of graphs based on their structure, providing a valuable tool for tasks such as graph categorization.
Supervisor: Jesús Pineda Examiner: Giovanni Volpe Opponent: Gideon Jägenstedt
Jesper Boberg and Anders Segerlund will defend their Master thesis in MPCAS at the Chalmers University of Technology on 14 June 2022 at 16:00.
Title: Early detection of rare evens: Predicting battery cell deviations.
Despite rigorous quality control in battery cell production, the production process is still subject to quality deviations. These quality deviations; known as “rare events” initially act as inherent passive quality deviations and may not affect the performance of the battery. Yet, a passive quality deviation can transition into an active quality deviation that give rise to behavioral deviations in the battery cell at some point during the battery’s lifetime. An active quality deviation may cause the entire battery to misbehave and eventually fail. This thesis investigates the possibility of predicting these cell deviations in car batteries. Better predictions of these events would avoid expensive and troublesome car failures and instead enable preventive car maintenance to solve the problem.
In this thesis different models have been created and evaluated with the aim of preventing these deviations. The dataset is supplied by Volvo Cars and contains a large amount of data collected from BEV cars where the arguably largest challenge comes from the imbalance of the dataset. In addition to the modelling, the thesis will include a thorough data analysis with the aim of improving both the dataset itself and the data collection process at Volvo Cars.
These deviations occur extremely rarely which also makes a relatively large amount of false positives difficult to avoid. The results show that a quite simple time series model can catch these deviations well but also brings along a large amount of false positives. A recurrent neural network was able to improve this significantly, still being able to catch the deviations while producing a lot fewer false positives.
Name of the master programme: MPALG – Computer Science: Algorithms, Languages, and Logic, MPCAS – Complex Adaptive Systems Examiner: Giovanni Volpe Supervisor: Herman Johnsson (Volvo Cars) Opponent: Jonathan Stålberg and Josef Gullholm5
David Rinman will defended his Master thesis in MPCAS at the Chalmers University of Technology on 13 June 2022 at 13:00.
Title: Monitoring Monitors; ML-based Anomaly Detection in Loudspeakers
Measuring the input voltage and current passing through a loudspeker and comparing the results to a parametric model is a way to monitor the condition of a loudspeaker in amplifiers. Prior research has shown that this can be done using music as input signal and can therefore work in commercial audio applications during normal operation. However, this solution requires modelling the specific loudspeaker setup which can be impractical in real-world scenarios. The aim of this project is to attempt to overcome these limitations by applying machine learning to the problem of anomaly detection in loudspeakers. Data is collected while playing music through two real functioning speakers where anomalies are simulated by disturbing the movement of their diaphragms. Three models are proposed and evaluated, two of which are based on deep neural networks. The results show that all three models are capable of learning a representation of one of the loudspeakers and detect deviances in these representations, however not for all of the simulated anomalies. Furthermore, the robustness of the models in prescence of nonlinear loudspeaker behavior is examined, and the limitations and benefits of the models are discussed. Finally, suggestions for future research directions are proposed.
Name of the master programme: MPCAS – Complex Adaptive Systems Examiner: Giovanni Volpe Supervisor: Jesper Pedersen (MusicTribe) Opponent: TBA
Kasper Hall and Noell Hall defended their Master thesis in MPCAS at the Chalmers University of Technology on 10 June 2022 at 14:00. Congrats!
Title: Interference Object Detection using TensorFlow Lite and Transfer Learning for Android Devices
With the rapid evolution of machine learning and artificial intelligence faster and more robust network architectures are developed. This is possible due to the increase in computational power, improved algorithms and the creation of large scale annotated datasets. Re-purposing these state of the art networks using transfer learning allows for customized models to be created and applied to niche problems. In this paper, we create an object detection application able to detect interference points in anechoic testing chambers. The application runs detection on a mobile device using networks created with TensorFlow Lite. Utilizing the detection result the application can give advice on how to improve the installation in the testing chamber and can thus enforce a baseline for how installations are conducted increasing the repeatability of tests.
Name of the master programme: MPCAS – Complex Adaptive Systems Examiner: Giovanni Volpe Supervisor: Giovanni Volpe and Christian Heina, Ericsson Opponent: Angelo Barona Balda
Angelo Barona Balda defended his Master thesis in MPCAS at the Chalmers University of Technology on 10 June 2022 at 13:00. Congrats!
Title: Playful Experiments with Macroscopic Active Matter
Active matter is a substance or system composed of individual agents that consume energy. These agents use the energy to vibrate, self-propel, or apply force to their surroundings.
As it is such a strong theoretical resource, professors use active matter as a gateway to introduce students into non-equilibrium research. However, it is difficult to intuitively explain the behavior of microscopic particles. To better visualize it, academics frequently use simulations , both in classrooms and in research.
In this study, we use the HEXBUGS to replicate active matter simulations performed in previous research papers. We create experiments that are didactic, understandable, and easy to reproduce. Through this, we prove that HEXBUGS behave like active particles.
Name of the master programme: MPCAS – Complex Adaptive Systems Examiner: Giovanni Volpe Supervisor: Giovanni Volpe, Aykut Argun Opponent: Noel Hall, Kasper Hall
Yanuar Rizki Pahlevi defended his Master thesis in MPCAS at the Chalmers University of Technology on 9 June 2022 at 17:00. Congrats!
Title: Deep Learning for Optical Tweezers. DeepCalib Implementation for Brownian Motion with Delayed Feedback
Brownian motion with delayed feedback theoretically studied to take control of Brownian particle movement’s direction. One can use optical tweezers to implement delayed feedback. Calibrating optical tweezers with delay implemented is not an easy job. In this study, Deep learning technique using Long Short Term Memory (LSTM) layer as main composition of the model to calibrate the trap stiffness and to measure the delayed feedback employed, using the trapped particle trajectory as an input. We demonstrate that this approach is outperforming variance methods in order to calibrate stiffness, also outperforming approximation method to measure the delay in harmonic trap case.
Name of the master programme: MPCAS – Complex Adaptive Systems Examiner: Giovanni Volpe Supervisor: Aykut Argun Opponent: Ivan Gentile Japiassu
Santhosh Shivan Gurumurthy defendeded his Master thesis in MPCAS at the Chalmers University of Technology on 2 June 2022 at 19:00. Congrats!
Title: Digital video microscopy using deep learning
Particle tracking is essential in various fields of science, from study of cells in biology to studying particle dynamics in physics. The standard classical methods are algorithmic in nature, requiring heavy manual work to tweak the parameter to various conditions and tasks. In this project, we try to apply four such methods to different particle tracking tasks and show that these are better alternatives to standard methods. We use Deeptrack as the base framework for all our tasks. We show that CNN can track a single Janus particle suspended in a liquid and outperforms a standard method called radial center method especially under noisy and unsymmetrical particle conditions. We then demonstrate how the LodeSTAR algorithm outperforms the YOLO algorithm in tracking the centers of multiple particles in an image. Subsequently, we apply a method called MAGIK (Motion Analysis through GNN Inductive Knowledge) to predict the trajectories of the particles in a video.
Name of the master programme: MPCAS – Complex Adaptive Systems Examiner: Giovanni Volpe Supervisor: Jesús Pineda Opponent: Emil Jansson
Emil Jansson defendeded his Master thesis in MPCAS at the Chalmers University of Technology on 2 June 2022 at 18:00. Congrats!
Title: Evolutionarily Emergent Foraging Strategies for Active Agents
Microbes, insects, birds, and mammals. Many forms of life depend on the search for food to survive. One search strategy that has been observed in nature is a levy flight, where an animal moves from area to area in long stretches to then explore the local environment. Levy flights can be described as statistical mathematical phenomena where the steps lengths of the agent’s movement follow a heavy tailed distribution. Earlier studies have shown that in certain environments, a middle ground between ballistic Levy flights and Brownian motion is more efficient than the outlier strategies. This thesis expands on those results by investigating which strategies perform best in an environment where local conditions change as one moves through space. We find that using a strategy that adapts to local conditions does not necessarily perform well if it does not consider the changing nature of the environment. We also let a neural network evolve using a genetic algorithm and let it optimize the movement of an agent which leads to efficient searches.
Name of the master programme: MPCAS – Complex Adaptive Systems Examiner: Giovanni Volpe Supervisor: Giovanni Volpe Opponent: Santhosh Shivan Gurumurthy
Isak Schwartz and William Åkvist defended their Master thesis in MPALG and MPCAS at the Chalmers University of Technology on 1 June 2022 at 16:00. Congrats!
Title: Active learning in deep convolutional neural networks for image segmentation
Evaluating data-centric approaches to improving performance in seat belt localization from images
The sitting position and seat belt orientation of passengers in automobiles can be crucial in the event of a collision. In order to warn passengers of unsafe positions, deep learning models in the form of neural networks can be used to identify the seat belt from image data. Performance of neural networks can be increased by improving the model (model-centric approaches) or by improving the data used to train the model (data-centric approaches).
In this thesis we compare the segmentation performance gains from model-centric approaches to data-centric approaches including stratified sampling, balancing, label error reduction and active learning. Active learning is the process of iteratively choosing data points for labeling according to the expected improvement in model performance. No new model architecture was found that improved performance, but the model training time was sped up by four times without performance loss. Stratified sampling, balancing and error reduction did not improve performance.
In active learning, images to be labeled were selected according to the model’s uncertainty. Several uncertainty metrics were used, all leading to an improvement when using active learning. The best result showed that we achieved 95% and 99% of the best baseline performance using 19% and 23% less data respectively.
Name of the master programme: MPALG – Computer Science: Algorithms, Languages, and Logic, MPCAS – Complex Adaptive Systems Examiner: Giovanni Volpe Supervisor: Tomas Björklund, Sheng Huang (Volvo Cars Corporation) Opponent: Rohini Bisht, Selomie Kindu