Time-series interpretability in healthcare presents unique challenges in understanding how deep learning models derive predictions from sequential medical data. This study evaluates four key explainable AI methodologies—backpropagation-based, perturbation- based, and counterfactual-based techniques—to determine their effectiveness for interpreting time-series models in healthcare. Through comprehensive analysis using ECG and EEG datasets, we identify the most suitable approaches for providing meaningful insights in healthcare applications while maintaining model accuracy and interpretability. Our evaluation encompasses multiple dimensions including complexity, faithfulness, reliability, and counterfactual analysis, providing a thorough framework for selecting appropriate interpretability methods based on specific healthcare contexts and requirements.
This study explores the impact of text prompt variations on CLIP’s visual understanding and classification accuracy. As a foundational zero-shot classifier, CLIP aligns image and text representations, but its sensitivity to subtle prompt changes raises concerns about robustness and reliability. By systematically varying prompt phrasing, including contextual and hallucination-based descriptions, we analyze their influence on object localization and attention alignment. Our methodology integrates GradCAM for visualization and evaluates performance using metrics such as AUC-ROC, PR-AUC, and IoU. Key findings reveal that enriched prompts enhance interpretability and attention concentration, improving object isolation in cluttered scenes. These results underscore the importance of effective prompt design in advancing CLIP’s real-world applicability.
Causal discovery methods have advanced significantly, yet their potential in manufacturing systems remains underutilized. This project investigates the applicability and performance of score- based causal discovery methods, including GES, NOTEARS, DAG- GNN, and GOLEM, in manufacturing contexts. Using the causalAssembly dataset, a semisynthetic data generator grounded in real-world manufacturing data, we benchmark these methods against traditional constraints-based and functional approaches. Our analysis evaluates their ability to uncover causal relationships in high-dimensional, noisy, and dynamic industrial data, focusing on key metrics such as Structural Hamming Distance (SHD) and F1 score. The results provide valuable insight into the strengths, limitations, and practical applicability of causal discovery methods, guiding their adaptation to tasks such as quality assurance, root cause analysis, and predictive maintenance in complex manufacturing environments.
Reinforcement Learning (RL) and Causal Discovery (CD) are key areas in artificial intelligence. RL helps optimize decisions using interaction data, while CD identifies cause-and-effect relationships. The Causality Aware Reinforcement Learning (CARL) framework combines these approaches to enhance decision-making and interpretability. This report reproduces the CARL framework, confirming its effectiveness in improving decision-making and model performance across different scenarios. The original CARL framework uses fixed exploration strategies and simple causal models, which limit its adaptability in complex environments. To address these issues, an enhancement called Dynamic Causal Exploration (DCE) is proposed. DCE adjusts explo- ration based on the accuracy of causal models, aiming to improve learning and adaptability. While the original CARL framework was tested with two simulated tasks, the Coffee and Taxi environments, this report focuses on reproducing these results and exploring the limitations. The proposed DCE method remains theoretical and was not fully tested in these tasks due to time constraints.
This project aims to create an interactive visualization tool to enhance the interpretability of complex machine learning models for non-technical stakeholders. Using local (LIME, SHAP) and global interpretability methods, our tool will provide clear, interactive explanations for individual model predictions and overall feature importance, enabling users to explore how specific features influence outcomes. This approach is designed to improve comprehension and trust in AI systems, especially in high-stakes fields like healthcare and finance. Experimental metrics will assess the effectiveness of different interpretability methods, contributing insights into the usability of XAI tools for responsible AI practice.
High performing deep neural networks in computer vision often function as black boxes, making fairness and explainability challenging due to limited insight into their decision-making processes. In this paper, we propose a method for bias detection in vision networks by leveraging automatic concept labeling of neurons. We present an approach to evaluate biases related to sensitive attributes including race, gender, and age by analyzing how these attributes are internally represented when predicting people’s professions from images. Our method enables the identification of biases within the model’s internal representations through explanations. Using datasets such as IdenProf and Meta FACET, we demonstrate the effectiveness of our approach in detecting biases, providing valuable insights that can lead to more equitable outcomes in real-world applications.
This project introduces Explainable Adaptive Weight Tuning (ExAWT), a framework designed to enhance the interpretability and performance of the language models. The study applies two complementary approaches: fine-tuning a baseline BERT model, where interpretability is evaluated using LIME, and implementing the ExAWT method, which utilizes gradient-based weighting to optimize layer-specific contributions during fine-tuning. Both approaches are evaluated on the BoolQ dataset from SuperGLUE, with the baseline showcasing enhanced task-specific accuracy and interpretability through explainability techniques, and the ExAWT method demonstrating efficiency and scalability by incorporating gradient insights directly into the fine-tuning process. Together, these approaches highlight innovative pathways for creating transparent, efficient, and robust language models tailored to domain-specific applications.
Congestive Heart Failure (CHF) is a chronic condition in which the heart’s ability to pump blood is compromised, leading to reduced oxygen supply to vital organs. Early detection of CHF is critical to improve patient outcomes and reduce healthcare costs. In this project, we aim to reimplement the convolutional neural network (CNN) approach for detecting congestive heart failure as outlined in the paper “A convolutional neural network approach to detect congestive heart failure.” With no publicly available code, we will reproduce the methodology by developing a CNN-based model that processes electrocardiogram (ECG) signals to classify CHF. Our focus will be on ensuring the model’s interpretability by employing an Explainable AI technique called GradCAM to provide insights into the factors driving predictions. This will enhance the model’s transparency and potential clinical applicability, improving trust in AI-assisted diagnostics.
Conditional independence (CI) testing is a fundamental problem in causal discovery, feature selection, machine learning and statistics, and this is particularly challenging in high-dimensional settings. In higher dimensions, there is increased data requirements and more non-linear relationships and we cannot use simple tests like chi- squared and t tests. Therefore, we use more complex sophisticated approaches like Hilbert-Schmidt independence criterion (HSIC) and Mutual information neural estimation (MINE). Here we have tried to implement and identify the shortcomings of MINE and tried to present a theoretical framework for modifying MINE for conditional independence testing. Though the results were not really convincing the theoretical framework behind is expected to be foolproof.