154092 research outputs found
Sort by
Do ImageNet-trained models learn shortcuts? The impact of frequency shortcuts on generalization
No full textFrequency shortcuts refer to specific frequency patterns that models heavily rely on for correct classification. Previous studies have shown that models trained on small image datasets often exploit such shortcuts, potentially impairing their generalization performance. However, existing methods for identifying frequency shortcuts require expensive computations and become impractical for analyzing models trained on large datasets. In this work, we propose the first approach to more efficiently analyze frequency shortcuts at a large scale. We show that both CNN and transformer models learn frequency shortcuts on ImageNet. We also expose that frequency shortcut solutions can yield good performance on out-of-distribution (OOD) test sets which largely retain texture information. However, these shortcuts, mostly aligned with texture patterns, hinder model generalization on rendition-based OOD test sets. These observations suggest that current OOD evaluations often overlook the impact of frequency shortcuts on model generalization. Future benchmarks could thus benefit from explicitly assessing and accounting for these shortcuts to build models that generalize across a broader range of OOD scenarios. Codes are available at https://github.com/nis-research/hfss
The Impact of Generative Artificial Intelligence Tools in Project-Based Learning
Full text linkWe investigate the potential risks and benefits for students utilising Generative Artificial Intelligence (GAI), specifically OpenAI’s ChatGPT and GitHub’s Copilot, to generate solutions instead of independently creating them in the traditional educational setting. The rapid advancements in GAI have transformed numerous domains, including software development and software engineering education. While these tools offer unprecedented convenience and efficiency, there are growing concerns regarding their potential implications for academic integrity and genuine student learning. We report on a pilot study in which 40 students, who completed a first-semester course on object-oriented programming, re-engage in a comparable programming project in limited time using GAI tools. In this pilot study, we aim to assess the extentto which students can rely on GAI tools to generate solutions for large programming assignments, to investigate the impact of AI-driven codegeneration on students’ understanding of fundamental programming concepts and problem-solving abilities, and to explore the perspectives ofeducators and students on the implications and long-term consequences of integrating AI-assisted coding tools in the learning process
Examining the market sample size for machine learning-based mass appraisal: a case study in Pendik district of Istanbul
No full textDetermining the optimum market sample size for Machine Learning (ML)-based modelling is important to achieve accurate, reliable and economical results in mass appraisal process. Insufficient market samples can lead to the model under-representing the diversity of the market, while many market samples may bring unnecessary computational costs. In this study, a methodological framework and comparison of model performances according to sample size for mass appraisal with ML was designed. A case study was performed in the densely populated Pendik district of Istanbul. 121 variables were determined in structural, spatial and local categories affecting the value of residential real estate, and the datasets were organised in the GIS environment. A total of 142 models were developed with increase of 500 market samples with the Random Forest algorithm, and models were evaluated with performance metrics. The modelling results showed that the models make estimations with high prediction accuracy and increasing the number of market samples improves performance in all metrics. When detecting the model performances based on the market sample size, the general distribution of the breakpoints was quite similar. It was concluded that high-accuracy results can be obtained with fewer market samples when the specific performance range is considered sufficient.</p
Accessibility to cancer medicines in Europe:Towards equitable access and fair pricing
No full textThis dissertation explores disparities in access to novel cancer medicines across Europe from the perspectives of patients, healthcare providers, and national authorities. It addresses challenges related to accessibility, pricing, and the socioeconomic impact of cancer, with the aim of informing evidence-based policies to support fair and sustainable cancer care.In the first part, access disparities were assessed through surveys with hospital pharmacists and oncology specialists across several European countries. Findings revealed significant variation in time to access and availability of novel cancer medicines, influenced by hospital type, country-specific regulations, and medicine type. Combination therapies were generally less accessible than monotherapies, and access was often faster in specialized hospitals and countries with strong early access programs. The second part of the dissertation examined medicine pricing. We analyzed managed entry agreements (MEAs), revealing differences in how countries negotiate and report discounts. Although MEAs aim to reduce costs, confidentiality limits transparency and comparability. We also presented confidential hospital-level pricing data from nine countries, showing wide disparities and emphasizing the need for standardized reporting to strengthen negotiations and collaboration.The third part investigated the socioeconomic impact of cancer from the patient perspective through the SEC study, involving over 3,000 patients in 25 countries. High levels of financial toxicity were reported, particularly among younger, lower-income, self-employed, and less-educated patients. A sub-study on adolescents and young adults (AYAs) found that 79% experienced financial difficulties, with limited support tailored to their needs. Many healthcare professionals remained unaware of these challenges.The dissertation concludes with six key recommendations, including harmonized EU regulations for early access, improved use of real-world evidence, discontinuation of external reference pricing, and enhanced price transparency. It also calls for the development of validated tools to identify vulnerable patients and inform targeted, supportive policies across Europe
Advancing patient-specific models of subthalamic deep brain stimulation
No full textDeep brain stimulation (DBS) in the subthalamic region is an established therapy for movement disorders like Parkinson’s disease (PD). Computational models of DBS provide insight into the electric field generated by DBS and the response it has on the neural tissue in the brain.Patient-specific models of DBS include the anatomical information of the patient to predict the effects of DBS. These models have been used to aid neurosurgeons with selecting an optimal DBS location and neurologists with the clinical programming of DBS to select therapeutic parameter settings.The subthalamic nucleus (STN) is one of the most common surgical targets for the treatment of PD. However, there is clinical debate if optimal therapeutic effects are achieved with stimulation within, or dorsal to, the STN. As such, observations of therapeutic benefit while stimulating white matter pathways instead of the grey matter nucleus have led to the emergence of a concept known as connectomic DBS modeling, which characterizes brain network connections that are modulated by stimulation. While patient-specific DBS models have been unsuccessful in defining a singular therapeutic target for PD, the introduction of connectomic modeling has shifted hypotheses on the control of symptoms from target areas within a specific nucleus to specific axonal pathways that course in/out or around the nucleus.This dissertation describes advances in patient-specific DBS computational modeling methodology seeking to identify a balance between biophysical realism and computational simplicity. It offers insights into DBS electric field modeling in the brain and introduces new softwaretools designed for clinical research on subthalamic DBS. The influence of electrode location and stimulation parameter selection on the axonal pathways is explored, as well as how stimulation sites can theoretically be designed to target specific pathways associated with symptom reduction. The work in this dissertation demonstrates the advancements made in patient-specific models of subthalamic DBS from the biophysical concept of connectomic DBS modeling to its therapeutic application in a prospective clinical trial
Advanced packaging for SiC power modules
No full textThe PhD research focuses on exploring solutions to the interconnection challenges in packaging that hinder the exploitation of the full potential of SiC semiconductor power modules. These packaging solutions aim to enable SiC power devices to operate at higher power densities, switching speeds, efficiency, and reliability. To be more specific, a systematic study of topside interconnection and substrate attachment regarding design, materials, process, simulation, testing, and failure analysis for SiC power modules is conducted
From remote monitoring to local action
No full textAs cities grow, more people face poor living conditions and greater wealth inequality. By overcoming conceptual, data-based and computational challenges, and focusing on cities in sub-Saharan Africa, a new study demonstrates how to measure this trend at scale
Planner3D:LLM-enhanced graph prior meets 3D indoor scene explicit regularization
No full textCompositional 3D scene synthesis has diverse applications across a spectrum of industries such as robotics, films, and video games, as it closely mirrors the complexity of real-world multi-object environments. Conventional works typically employ shape retrieval based frameworks which naturally suffer from limited shape diversity. Recent progresses have been made in object shape generation with generative models such as diffusion models, which increases the shape fidelity. However, these approaches separately treat 3D shape generation and layout generation. The synthesized scenes are usually hampered by layout collision, which suggests that the scene-level fidelity is still under-explored. In this paper, we aim at generating realistic and reasonable 3D indoor scenes from scene graph. To enrich the priors of the given scene graph inputs, large language model is utilized to aggregate the global-wise features with local node-wise and edge-wise features. With a unified graph encoder, graph features are extracted to guide joint layout-shape generation. Additional regularization is introduced to explicitly constrain the produced 3D layouts. Benchmarked on the SG-FRONT dataset, our method achieves better 3D scene synthesis, especially in terms of scene-level fidelity. The source code will be released after publication. 3D indoor scene synthesis, generative model, scene graph, large language model, spatial arrangement, latent diffusion.</p
Learning to Harmonize Cross-Vendor X-ray Images by Non-linear Image Dynamics Correction
No full textIn this paper, we explore how conventional image enhancement can improve model robustness in medical image analysis. By applying commonly used normalization methods to images from various vendors and studying their influence on model generalization in transfer learning, we show that the nonlinear characteristics of domain-specific image dynamics cannot be addressed by simple linear transforms. To tackle this issue, we reformulate the image harmonization task as an exposure correction problem and propose a method termed Global Deep Curve Estimation (GDCE) to reduce domain-specific exposure mismatch. GDCE performs enhancement via a pre-defined polynomial function and is trained with a “domain discriminator”, aiming to improve model transparency in downstream tasks compared to existing black-box methods. Code available at https://github.com/YCL92/GDCE.</p
Monthly rainfall forecasting using deep learning methods in big databases:a case study of northwestern Iran
No full textThis investigation introduces a monthly rainfall forecast model employing Deep Belief Network (DBN) techniques for rainfall prediction in northwestern Iran. Meteorological, climatic, and topographic data were used to model the rainfall. Since the parameter combination directly impacts the accuracy of results, an optimized combination of input parameters was selected through the Genetic Algorithm (GA). Moreover, considering the imbalanced nature of the dataset, Random Oversampling and Random Undersampling were exploited to increase the model's performance. The predictive capabilities of the suggested model were compared with those of a Multilayer Perceptron (MLP) and Long Short-Term Memory (LSTM) methods. The findings revealed that the MLP achieved an accuracy of 0.67, while the DBN exhibited a higher accuracy of 0.71. The comparison proved the superiority of the developed model based on DBN. The result also revealed that the Random Undersampling technique successfully identified classes with less distribution.</p