As an able-bodied human, understanding what someone says during a phone call with a lot of background noise is usually a task that is quite easy for us as we are aware of what the information is we want to hear, e.g. the voice of the person we are talking to, and the information that is noise, e.g. music or ambient noise in the background. While dealing with noise of all kinds for most humans proves to be the easiest, it is a very hard task for algorithms to deal with noisy data. Unfortunately for some beneficial and interesting applications, like Brain Computer Interfaces (short BCI) based on Electroencephalography (short EEG) data, noise is a very prevalent problem that greatly hinders the progress of making BCIs for real-life applications. In this thesis, we investigate what effect noise added to EEG data has on the classification accuracy of one Spiking Neural Network and one Convolutional Neural Network based classifier for a motor imagery classification task. The thesis shows that already relatively small amounts of noise (10% of original data) can have strong effects on the classification accuracy of the chosen classifiers. It also provides evidence that SNN based models have a more stable classification accuracy for low amounts of noise. Still, their classification accuracy after that declines more rapidly, while CNN based classifiers show a more linear decline in classification accuracy / Att förstå vad någon säger under ett telefonsamtal med mycket bakgrundsljud är en relativt enkel uppgift för en människa eftersom vi är duktiga på att urskilja vilken del av ljudet som är relevant, t.ex. rösten hos den vi pratar med, och vilken del av ljudet som är bakgrundsbrus, t.ex. musik eller omgivningsljud. Även om det är en enkel uppgift för en människa att filtrera bort olika sorters brus så är det betydligt svårare för en algoritm att hantera brusig data. Tyvärr finns det flertalet användbara och intressanta applikationsområden där svårigheten med brus orsakar betydande problem. Ett sådant exempel är braincomputer interfaces (BCI) baserade på elektroencefalografi (EEG) där brus är ett så pass utbrett problem att det begränsar möjligheten att använda BCI i verkliga tillämpningar. I detta examensarbete undersöks hur tillägget av brus till EEG-data påverkar noggrannheten på klassificeringen av hjärnaktivitet vid visualisering av olika rörelser. För detta ändamål jämfördes två typer av klassificerare: ett spiking neural network (SNN) och ett convolutional neural network (CNN). Examensarbetet visar att redan relativt små tillägg av brus (10%) kan ha stor påverkan på klassificeringens noggrannhet. Studien påvisar även att SNN-baserade modeller har en mer stabil noggrannhet för låga mängder brus, men att noggrannheten försämras snabbare vid ökad mängd brus än för CNN-baserade klassificerare som visar en mer linjär försämring.

The increasing popularity of drones has made it convenient to capture a large number of images of a property, which can then be used to build a 3D model. The conditions of buildings can be analyzed to plan renovations. This creates an interest for automatically identifying building materials, a task well suited for machine learning. With access to drone imagery of buildings as well as depth maps and normal maps, we created a dataset for semantic segmentation. Two different convolutional neural networks were trained and evaluated, to see how well they perform material segmentation. DeepLabv3+, which uses RGB data, was compared to Depth-Aware CNN, which uses RGB-D data. Our experiments showed that DeepLabv3+ achieved higher mean intersection over union. To investigate if the information in the depth maps and normal maps could give a performance boost, we conducted experiments with an encoding we call HMN - horizontal disparity, magnitude of normal with ground, normal parallel with gravity. This three channel encoding was used to jointly train two CNNs, one with RGB and one with HMN, and then sum their predictions. This led to improved results for both DeepLabv3+ and Depth-Aware CNN. / Den ökade populariteten av drönare har gjort det smidigt att ta ett stort antal bilder av en fastighet, och sedan skapa en 3D-modell. Skicket hos en byggnad kan enkelt analyseras och renoveringar planeras. Det är då av intresse att automatiskt kunna identifiera byggnadsmaterial, en uppgift som lämpar sig väl för maskininlärning.  Med tillgång till såväl drönarbilder av byggnader som djupkartor och normalkartor har vi skapat ett dataset för semantisk segmentering. Två olika faltande neuronnät har tränats och utvärderats för att se hur väl de fungerar för materialigenkänning. DeepLabv3+ som använder sig av RGB-data har jämförts med Depth-Aware CNN som använder RGB-D-data och våra experiment visar att DeepLabv3+ får högre mean intersection over union. För att undersöka om resultaten kan förbättras med hjälp av datat i djupkartor och normalkartor har vi kodat samman informationen till vad vi valt att benämna HMN - horisontell disparitet, magnitud av normalen parallell med marken, normal i gravitationsriktningen. Denna trekanalsinput kan användas för att träna ett extra CNN samtidigt som man tränar med RGB-bilder, och sedan summera båda predikteringarna. Våra experiment visar att detta leder till bättre segmenteringar för både DeepLabv3+ och Depth-Aware CNN.

In this project, a model based on a convolutional neural network have been developed with the aim of imputing missing genotype data. This model was based on an already existing autoencoder that was modified into a U-Net structure. The network was trained and used iteratively with the intention that the result would improve in each iteration. In order to do this, the output of the model was used as the input in the next iteration. The data used in this project was diploid genotype data, which was phased into haploids and then run separately through the network. In each iteration, the new haploids were generated based on the output haploids. These were used as in input in the next iteration. The result showed that the accuracy of the imputation improved slightly in every iteration. However, it did not surpass the same model that was trained for one single iteration. Further work is needed to make the model more useful.

Interpretation of electroencephalograms (EEGs) is a process that is still dependent on the subjective analysis of the examiner. The interrater agreement, even for relevant clinical events such as seizures, can be low. For instance, the differences between interictal, ictal, and post-ictal EEGs can be quite subtle. Before making such low-level interpretations of the signals, neurologists often classify EEG signals as either normal or abnormal. Even though the characteristics of a normal EEG are well defined, there are some factors, such as benign variants, that complicate this decision. However, neurologists can make this classification accurately by only examining the initial portion of the signal. Therefore, in this thesis, we explore the hypothesis that high performance machine classification of an EEG signal as abnormal can approach human performance using only the first few minutes of an EEG recording. The goal of this thesis is to establish a baseline for automated classification of abnormal adult EEGs using state of the art machine learning algorithms and a big data resource – The TUH EEG Corpus. A demographically balanced subset of the corpus was used to evaluate performance of the systems. The data was partitioned into a training set (1,387 normal and 1,398 abnormal files), and an evaluation set (150 normal and 130 abnormal files). A system based on hidden Markov Models (HMMs) achieved an error rate of 26.1%. The addition of a Stacked Denoising Autoencoder (SdA) post-processing step (HMM-SdA) further decreased the error rate to 24.6%. The overall best result (21.2% error rate) was achieved by a deep learning system that combined a Convolutional Neural Network and a Multilayer Perceptron (CNN-MLP). Even though the performance of our algorithm still lags human performance, which approaches a 1% error rate for this task, we have established an experimental paradigm that can be used to explore this application and have demonstrated a promising baseline using state of the art deep learning technology. / Electrical and Computer Engineering

Application of deep learning in radiology has the potential to automate workflows, support radiologists with decision support, and provide patients a logic-based algorithmic assessment. Unfortunately, medical datasets are often not uniformly distributed due to a naturally occurring imbalance. For this research, a multi-classification of liver MRI sequences for imaging of hepatocellular carcinoma (HCC) was conducted on a highly imbalanced clinical dataset using deep convolutional neural network. We have compared four multi classification classifiers which were Model A and Model B (both trained using imbalanced training data), Model C (trained using augmented training images) and Model D (trained using under sampled training images). Data augmentation such as 45-degree rotation, horizontal and vertical flip and random under sampling were performed to tackle class imbalance. HCC, the third most common cause of cancer-related mortality [1], can be diagnosed with high specificity using Magnetic Resonance Imaging (MRI) with the Liver Imaging Reporting and Data System (LI-RADS). Each individual MRI sequence reveals different characteristics that are useful to determine likelihood of HCC. We developed a deep convolutional neural network for the multi-classification of imbalanced MRI sequences that will aid when building a model to apply LI-RADS to diagnose HCC. Radiologists use these MRI sequences to help them identify specific LI-RADS features, it helps automate some of the LIRADS process, and further applications of machine learning to LI-RADS will likely depend on automatic sequence classification as a first step. Our study included an imbalanced dataset of 193,868 images containing 10 MRI sequences: in- phase (IP) chemical shift imaging, out-phase (OOP) chemical shift imaging, T1-weighted post contrast imaging (C+, C-, C-C+), fat suppressed T2 weighted imaging (T2FS), T2 weighted imaging, Diffusion Weighted Imaging (DWI), Apparent Diffusion Coefficient map (ADC) and In phase/Out of phase (IPOOP) imaging. Model performance for Models A, B, C and D provided a macro average F1 score of 0.97, 0.96, 0.95 and 0.93 respectively. Model A showed higher classification scores than models trained using data augmentation and under sampling. / Thesis / Master of Science (MSc)

The thesis explores the application of Convolutional Neural Networks (CNNs) in the fashion industry, focusing on fashion style classification and image retrieval. Employing transfer learning, the study investigates the effectiveness of fine-tuning pre-trained CNN models to adapt them for a specific fashion recognition task by initially performing an extensive hyperparameter optimisation, utilising the Optuna framework.  The impact of dataset size on model performance was examined by comparing the accuracy of models trained on datasets containing 2000 and 8000 images. Results indicate that larger datasets significantly improve model performance, particularly for more complex models like EfficientNetV2S, which showed the best overall performance with an accuracy of 85.38% on the larger dataset after fine-tuning. The best-performing and fine-tuned model was subsequently used for image retrieval as features were extracted from the last convolutional layer. These features were used in a cosine similarity measure to rank images by their similarity to a query image. This technique achieved a mean average precision (mAP) of 0.4525, indicating that CNNs hold promise for enhancing fashion retrieval systems, although further improvements and validations are necessary. Overall, this research highlights the versatility of CNNs in interpreting and categorizing complex visual data. The importance of well-prepared, targeted data and refined model training strategies is highlighted to enhance the accuracy and applicability of AI in diverse fields.

As autonomous vehicles (AVs) evolve, ensuring their safety extends beyond traditional met- rics. While current nominal safety scores focus on the timeliness of AV responses like latency or instantaneous response time, this paper proposes expanding the concept to include spatial configurations formed by obstacles with respect to the ego-vehicle. By analyzing these spatial relationships, including proximity, density and arrangement, this research aims to demon- strate how these factors influence the safety force field around the AV. The goal is to show that beyond meeting Responsibility-Sensitive Safety (RSS) metrics, spatial configurations significantly impact the safety force field, particularly affecting path planning capability. High spatial occupancy of obstacle configurations can impede easy maneuverability, thus challenging safety-critical modules like path planning. This paper aims to capture this by proposing a safety score that leverages the ability of modern computer vision techniques, par- ticularly image segmentation models, to capture high and low levels of spatial and contextual information. By enhancing the scope of nominal safety to include such spatial analysis, this research aims to broaden the understanding of drivable space and enable AV designers to evaluate path planning algorithms based on spatial configuration centric safety levels. / Master of Science / As self-driving cars become more common, ensuring their safety is crucial. While current safety measures focus on how quickly these cars can react to dangers, this paper suggests that understanding the spatial relationships between the car and obstacles is just as important, and needs to be explored further. Prior metrics use velocity and acceleration of all the actors, to determine the safe-distance of obstacles from the vehicle, and determine how fast the car should react before a predicted collision. This paper aims to extend the scope of how safety is viewed during normal operating conditions of the vehicle by considering the arrangement of obstacles around it as an influencing factor to safety. By using advanced computer vision techniques, particularly models that can understand images in detail, this research proposes a new spatial safety metric. This score considers how well the car navigates through dense environments by understanding the spatial configurations that obstacles form. By studying these factors, I wish to introduce a metric that improves how self-driving cars are designed to navigate and path plan safely on the roads.

Computational psychiatry describes psychological phenomena as abnormalities in biological computations. Current available technologies span multiple organizational and temporal domains, but there remains a knowledge gap with respect to neuromodulator dynamics in humans. Recent efforts by members of the Montague Laboratory and collaborators adapted fast scan cyclic voltammetry (FSCV) from rodent experiments for use in human patients already receiving brain surgery. The process of modifying established FSCV methods for clinical application has led improved model building strategies, and a new "random burst" sensing protocol. The advent of random burst sensing raises questions about the capabilities of in-vivo electrochemistry techniques, while opening introducing possibilities for novel approaches. Through a series of in-vitro experiments, this study aims to explore and validate novel electrochemical sensing approaches. Initial expository experiments tested assumptions about waveform design to detect dopamine concentrations by reducing amplitude and duration of forcing functions, as well as distinguishing norepinephrine concentrations. Next, large data sets collected on mixtures of dopamine, serotonin and pH validated a newly proposed "low amplitude random burst sensing" protocol, for both within-probe and out-of-probe modeling. Data collected on the same set of solutions also attempted to establish an order-millisecond random burst sensing approach. Preliminary endeavors into using convolutional neural networks also provided an example of an alternative modeling strategy. The results of this work challenge existing assumptions of neurochemistry, while demonstrating the capabilities of new neurochemical sensing approaches. This study will also act as a springboard for emerging technological developments in human neurochemistry. / Doctor of Philosophy / Neuroscience characterizes nervous system functions from the cellular to the systems level. A gap in available technologies has prevented neuroscientist from studying how changes in the molecular dynamics in the brain relate to psychiatric conditions. Recent efforts by the Montague Laboratory have adapted neurochemistry techniques for use in human patients. Consequently, a new "random burst sensing" approach was developed that challenged existing assumptions about electrochemistry. In this study, in-vivo experiments were conducted to push the limits of electrochemical sensing by reducing the voltage amplitude range and increasing sensing temporal resolution of electrochemical sensing beyond previously established limits. The results of this study offer novel neurochemistry approaches and act as a jumping off point for future technological developments.

Computer vision techniques are becoming increasingly popular. For example, face recognition is used to help police find criminals, vehicle detection is used to prevent drivers from serious traffic accidents, and written word recognition is used to convert written words into printed words. With the rapid development of vehicle detection given the use of deep learning techniques, there are still concerns about the performance of state-of-the-art vehicle detection techniques. For example, state-of-the-art vehicle detectors are restricted by the large variation of scales. People working on vehicle detection are developing techniques to solve this problem. This thesis proposes an advanced vehicle detection model, adopting one of the classical neural networks, which are the residual neural network and the region proposal network. The model utilizes the residual neural network as a feature extractor and the region proposal network to detect the potential objects' information. / Master of Science / Computer vision techniques are becoming increasingly popular. For example, face recognition is used to help police find criminals, vehicle detection is used to prevent drivers from serious traffic accidents, and written word recognition is used to convert written words into printed words. With the rapid development of vehicle detection given the use of deep learning techniques, there are still concerns about the performance of state-of-the art vehicle detection techniques. For example, state-of-the-art vehicle detectors are restricted by the large variation of scales. People working on vehicle detection are developing techniques to solve this problem. This thesis proposes an advanced vehicle detection model, utilizing deep learning techniques to detect the potential objects’ information.

Historical photographs can generate significant cultural and economic value, but often their subjects go unidentified. However, if analyzed correctly, visual clues in these photographs can open up new directions in identifying unknown subjects. For example, many 19th century photographs contain painted backdrops that can be mapped to a specific photographer or location, but this research process is often manual, time-consuming, and unsuccessful. AI-based computer vision algorithms could be used to automatically identify painted backdrops or photographers or cluster photos with similar backdrops in order to aid researchers. However, it is unknown which computer vision algorithms are feasible for painted backdrop identification or which techniques work better than others. We present three studies evaluating four different types of image embeddings – Inception, CLIP, MAE, and pHash – across a variety of metrics and techniques. We find that a workflow using CLIP embeddings combined with a background classifier and simulated user feedback performs best. We also discuss implications for human-AI collaboration in visual analysis and new possibilities for digital humanities scholarship. / Master of Science / Historical photographs can generate significant cultural and economic value, but often their subjects go unidentified. However, if these photographs are analyzed correctly, clues in these photographs can open up new directions in identifying unknown subjects. For example, many 19th century photographs contain painted backdrops that can be mapped to a specific photographer or location, but this research process is often manual, time-consuming, and unsuccessful. Artificial Intelligence-based computer vision techniques could be used to automatically identify painted backdrops or photographers or group together photos with similar backdrops in order to aid researchers. However, it is unknown which computer vision techniques are feasible for painted backdrop identification or which techniques work better than others. We present three studies comparing four different types of computer vision techniques – Inception, CLIP, MAE, and pHash – across a variety of metrics. We find that a workflow that combines the CLIP computer vision technique, software that automatically classifies photo backgrounds, and simulated human feedback performs best. We also discuss implications for collaboration between humans and AI for analyzing images and new possibilities for academic research combining technology and history.