This study aims to use purely data-driven causal discovery algorithms to assess and compare physics-based causal links in the field of climate science. Our dataset involves various measured and latent, non-linear and non-gaussian variables related to aerosol-cloud interactions. Our objective is to investigate unaccounted-for links that were previously unidentified and assess domain assumed links that may not be detected when applying causal discovery techniques to the raw dataset. Further- more, we seek to delineate the discoverable aspects and identify the limitations inherent in these causal discovery algorithms with simulated data with a known ground truth. We use four different causal discovery algorithms that use differ- ent methodologies for causal discovery including structural equation based, deep learning based, conditional independence based, and score based methods. This exploration may lead to potential modifications and enhancements to better tailor these algorithms for non-linear, noisy datasets commonly encountered in climate datasets, as well as prove to be useful for further in-depth analysis in this field.
This project delves into the intriguing phenomenon of double descent in machine learning, specifically within neural networks, utilizing the MNIST dataset. We aim to investigate how these factors influence model performance by manipulating both data points and model complexity. Employing SHAP, LIME and Saliency Maps as our analytical tools, we elucidate the interpretability and reliability of neural networks at various points along the double descent curve. Our hypothesis suggests that when the number of parameters (d) is close to the feature vector size (n), the performance reduces due to overfitting. By examining model explanations, we want to determine and explore whether the predictions are meaningful or driven by spurious correlations at various points of the double descent. We also want to look at various different analytical tools and compare them qualitatively. This project aims to provide insights into the relationship between model complexity and generalization, with potential implications for developing more interpretable and reliable neural networks in real-world applications.
In the era of artificial intelligence, multimodal models have emerged as a pivotal technology for interpreting complex data across various domains, including healthcare, autonomous navigation, and content recommendation. These models process and analyze multiple forms of data—text, images, audio, and video to make predictions or decisions. Despite their vast potential, a significant challenge persists: the explainability of these models. Our project addresses this challenge by enhancing the transparency and understandability of multimodal models. We suggest an approach that integrates both tabular and textual data in a unified multi- modal framework, leveraging a custom masking strategy for explainability. This technique not only provides deeper insights into how different data modalities influence model predictions but also overcomes the limitations of current state-of- the-art explainability methods, which often focus on single modalities.
Inspired by the work of Becker et al[ 2], this project aims to enhance the inter- pretability of deep neural network decisions in the audio domain. Despite the high accuracy of deep learning models in audio classification, their “black box” nature poses significant challenges in understanding their decision-making processes. This is particularly problematic in critical applications where trust and transparency are paramount. We leverage Layer-wise Relevance Propagation technique to identify significant features by two deep neural networks, namely AlexNet and AudioNet that process audio data in the forms of Spectrograms and Waveforms, respectively. We apply these models for three tasks: Sex, Digit Spoken and Accent classification on the AudioMNIST dataset. Along with traditional visual explanations, we have explored audible explanations using audible heatmaps. We assessed the effective- ness of LRP using systematic input data perturbation. By introducing audible explanations and comparing them with traditional visual explanations, we aim to significantly improve the interpretability of audio classification models, making their decisions more accessible and trustworthy to humans.
There have been significant advancements in relation extraction (RE) that have resulted in impressive benchmark accuracy. However, there is still significantly more to be explored when it comes to understanding the learning preferences in RE. In this work, we start by discussing the use of adversarial attacks to explore the model’s learning preference and robustness. Following this analysis, we implement a different counterfactual calibration method to adjust the prediction results of the RE model during testing thus reducing the over-dependency issue. This involves calculating the counterfactual prediction results by excluding the component that the model heavily relies on. Extensive experiments in multiple datasets indicate the necessity of introducing tuning-based methods in RE to further address the over-dependency issue.
Sentiment analysis, a critical task in Natural Language Processing (NLP), is em- ployed to classify tweets related to stocks by their positive or negative sentiments. While deep learning-based models, particularly BERT, achieve high accuracy in this domain, their inherent black-box nature hinders the understanding of how sentiment classification is performed. In this study, we apply BERT-based models to a dataset of stock-related tweets and explore their interpretability using Local Interpretable Model-agnostic Explanation (LIME) and SHapley Additive exPlana- tions (SHAP). By analyzing the attention weights and feature importance scores, we aim to shed light on the decision-making process of BERT models, especially in financial sentiment analysis. Additionally, we compare the performance of BERT with rule-based models such as Logistic Regression and Naive Bayes which offers transparency but sacrifice accuracy.
This project aims to develop advanced counterfactual explanation algorithms for behavioral and text data to enhance the transparency and accountability of AI systems, addressing the growing need for understandable AI decisions in critical sectors such as healthcare, finance, and human resources. This endeavor is pivotal as it directly responds to the challenge of making complex AI models explainable to non-expert users, thereby fostering trust and facilitating ethical AI practices. The task is inherently difficult due to the complexity of underlying AI models, the nuanced nature of behavioral and text data, and the challenge of generating explana- tions that are both accurate and easily comprehensible to humans. Existing methods often struggle to balance these aspects, particularly in providing actionable insights that are directly relevant to the users’ context. To overcome these challenges, our approach integrates cutting-edge techniques in natural language processing, causal inference, and optimization to create more precise, context-aware counterfactual explanations. By refining and extending current methodologies, the project aims to offer novel contributions that not only improve the clarity and relevance of expla- nations but also enhance the ability of AI systems to be scrutinized and improved, ensuring that they align more closely with human values and regulatory standards.
Predictive problems across various domains involve decision-making processes that rely on machine learning models built from relational data spread across mul- tiple tables. However, constructing machine learning models faces challenges in operating seamlessly across these relational structures, often requiring laborious manual processes of data joining and aggregation. To address these limitations, a Relational Deep Learning approach directly learns from data spread across multiple tables in data warehouses by viewing them as a heterogeneous graph. This transfor- mation serves as the basis for developing Graph Neural Network (GNN) predictive models. Given the inherent complexity of GNNs, the integration of explainability models becomes crucial, providing transparent insights into their decision-making processes. This project underscores the pivotal role of explainability in enhancing trust, identifying errors, and facilitating continuous improvement, ensuring that GNNs are not only accurate but also interpretable for real-world applications and regulatory compliance. Here, we will construct a GNN on the heterogeneous graph for a regression task and explain the predictions made by the model by attributing the prediction to its input features using four attribution methods namely Integrated Gradients, Saliency, Deconvolution, and Guided Backpropagation.
Recommendation systems have become an integral part of various online platforms, enhancing user experience by providing personalized content. However, there is a growing need for more transparent and interpret able recommendation systems to address user trust issues. This project focuses on counterfactual reasoning for explainable recommendation. The design of evaluation metrics is from two viewpoints: 1) user’s perspective and 2) model’s perspective. In this project we will experiment with different Large Language Models (LLMs) both open source and proprietary(till no cost limit) like GPT3.5, Llama2, Mistral to extract aspects and sentiments, and further compare it with the approach used by the paper (Sentires).
In response to the widespread integration of machine learning models in sensitive domains, our project addresses the need for interpretability in machine learning models. Leveraging post-hoc model-independent explanation techniques like LIME and SHAP, we aim to address the demonstrated vulnerability of these methods to deception, which effectively conceals inherent biases. If left unaddressed, these biases could erode trust in machine learning models, potentially resulting in misguided decisions with significant consequences. In this project, we explore the initial idea behind fooling LIME and SHAP, analyze existing techniques that are being used to counter it. We also propose using CT-GAN as a data generating technique to augment current methods of “Unfooling”.
This research project introduces a novel Large Language Model (LLM) framework designed to operate on locally stored user data, enabling personalized information retrieval and question-answering with direct citations from the user’s own datasets. Unlike traditional cloud-based LLMs, our approach ensures data privacy and rele- vance by processing information directly on the user’s device, mitigating concerns related to data security and internet dependency. By harnessing the capabilities of advanced natural language processing and machine learning algorithms, the proposed LLM dynamically generates accurate, contextually relevant responses with citations, enhancing the reliability and traceability of information. This local execution model not only addresses the challenge of hallucination in text generation but also significantly improves the utility of LLMs for users with specialized data needs or those working in privacy-sensitive environments. The outcome of this research paves the way for a new generation of LLM applications, where data sovereignty and personalized knowledge extraction are paramount.
This project aims to compare automated debiasing pipelines and qualitative debias- ing approaches for unstructured textual data, focusing on hate speech detection in online platforms. The challenge lies in the complexity of debiasing, particularly in handling linguistic variations and dialects, such as African American Vernacular English (AAVE), which are often misrepresented or mislabeled in training data, leading to biased outcomes. By leveraging novel debiasing techniques informed by subgroup linguistic characteristics, and using diverse data sources like Twitter and specialized datasets for underrepresented dialects, we plan to enhance the accuracy and fairness of hate speech detection systems. This contribution addresses the pressing need for adaptable and ethical debiasing methods that can keep pace with the evolving landscape of online discourse.
In this project, we propose different methodologies to integrate fairness metrics into machine learning pipelines, addressing the challenge of biased data. This is crucial as data-driven decisions increasingly impact society and making this integration is challenging due to the complex nature of fairness in data. The methodology involves directly integrating fairness constraints into the existing data- cleaning/ model-training frameworks like AutoML, AlphaClean and XGBoost. The experiments validate this approach on biased datasets like UCI Statlog (German Credit) and UCI Adult, aiming to balance data quality and fairness, contributing significantly to responsible AI development.
In this project, we explore how to improve machine learning (ML) models by cleaning and filling in missing data using different methods: mean, KNN, iterative imputation, and an advanced tool called AlphaClean. We want to see which method works best because good data quality is key to making accurate and fair predictions with ML. This challenge is tricky because every dataset has its own unique problems, and there’s no one-size-fits-all solution for fixing errors or filling gaps in the data. Our approach tests these methods by deliberately adding errors to our data, then cleaning it up and seeing how well the ML models perform afterward. We found out that while all methods have their strengths, understanding your specific data’s needs is most important. Our work helps show which data preparation methods might be best for different kinds of datasets, making it easier for people working with ML to get reliable and fair results from their models.