In today’s world, artificial intelligence (AI) is increasingly becoming a part of our daily lives. For this integration to be successful, it’s essential that AI systems can effectively interact with humans. This means making the AI system’s behavior more understandable to users and allowing users to customize the system’s behavior to match their preferences. However, there are significant challenges associated with achieving this goal. One major challenge is that modern AI systems, which have shown great success, often make decisions based on learned representations. These representations, often acquired through deep learning techniques, are typically inscrutable to the users inhibiting explainability and customizability of the system. Additionally, since each user may have unique preferences and expertise, the interaction process must be tailored to each individual. This thesis addresses these challenges that arise in human-AI interaction scenarios, especially in cases where the AI system is tasked with solving sequential decision-making problems. This is achieved by introducing a framework that uses a symbolic interface to facilitate communication between humans and AI agents. This shared vocabulary acts as a bridge, enabling the AI agent to provide explanations in terms that are easy for humans to understand and allowing users to express their preferences using this common language. To address the need for personalization, the framework provides mechanisms that allow users to expand this shared vocabulary, enabling them to express their unique preferences effectively. Moreover, the AI systems are designed to take into account the user’s background knowledge when generating explanations tailored to their specific needs.

In the age of artificial intelligence, Machine Learning (ML) has become a pervasive force, impacting countless aspects of our lives. As ML’s influence expands, concerns about its reliability and trustworthiness have intensified, with security and robustness emerging as significant challenges. For instance, it has been demonstrated that slight perturbations to a stop sign can cause ML classifiers to misidentify it as a speed limit sign, raising concerns about whether ML algorithms are suitable for real-world deployments. To tackle these issues, Responsible Machine Learning (Responsible ML) has emerged with a clear mission: to develop secure and robust ML algorithms. This dissertation aims to develop Responsible Machine Learning algorithms under real-world constraints. Specifically, recognizing the role of adversarial attacks in exposing security vulnerabilities and robustifying the ML methods, it lays down the foundation of Responsible ML by outlining a novel taxonomy of adversarial attacks within real-world settings, categorizing them into black-box target-specific, and target-agnostic attacks. Subsequently, it proposes potent adversarial attacks in each category, aiming to obtain effectiveness and efficiency. Transcending conventional boundaries, it then introduces the notion of causality into Responsible ML (a.k.a., Causal Responsible ML), presenting the causal adversarial attack. This represents the first principled framework to explain the transferability of adversarial attacks to unknown models by identifying their common source of vulnerabilities, thereby exposing the pinnacle of threat and vulnerability: conducting successful attacks on any model with no prior knowledge. Finally, acknowledging the surge of Generative AI, this dissertation explores Responsible ML for Generative AI. It introduces a novel adversarial attack that unveils their adversarial vulnerabilities and devises a strong defense mechanism to bolster the models’ robustness against potential attacks.

Social media platforms have become widely used for open communication, yet their lack of moderation has led to the proliferation of harmful content, including hate speech. Manual monitoring of such vast amounts of user-generated data is impractical, thus necessitating automated hate speech detection methods. Pre-trained language models have been proven to possess strong base capabilities, which not only excel at in-distribution language modeling but also show powerful abilities in out-of-distribution language modeling, transfer learning and few-shot learning. However, these models operate as complex function approximators, mapping input text to a hate speech classification, without providing any insights into the reasoning behind their predictions. Hence, existing methods often lack transparency, hindering their effectiveness, particularly in sensitive content moderation contexts. Recent efforts have been made to integrate their capabilities with large language models like ChatGPT and Llama2, which exhibit reasoning capabilities and broad knowledge utilization. This thesis explores leveraging the reasoning abilities of large language models to enhance the interpretability of hate speech detection. A novel framework is proposed that utilizes state-of-the-art Large Language Models (LLMs) to extract interpretable rationales from input text, highlighting key phrases or sentences relevant to hate speech classification. By incorporating these rationale features into a hate speech classifier, the framework inherently provides transparent and interpretable results. This approach combines the language understanding prowess of LLMs with the discriminative power of advanced hate speech classifiers, offering a promising solution to the challenge of interpreting automated hate speech detection models.

Large Language Models (LLMs) have displayed impressive capabilities in handling tasks that require few demonstration examples, making them effective few-shot learn- ers. Despite their potential, LLMs face challenges when it comes to addressing com- plex real-world tasks that involve multiple modalities or reasoning steps. For example, predicting cancer patients’ survival period based on clinical data, cell slides, and ge- nomics poses significant logistical complexities. Although several approaches have been proposed to tackle these challenges, they often fall short in achieving promising performance due to their inability to consider all modalities simultaneously or account for missing modalities, variations in modalities, and the integration of multi-modal data, ultimately compromising their effectiveness.This thesis proposes a novel approach for multi-modal tumor survival prediction to address these limitations. Taking inspiration from recent advancements in LLMs, particularly Mixture of Experts (MoE)-based models, a graph-guided MoE framework is introduced. This framework utilizes a graph structure to manage the predictions effectively and combines multiple models to enhance predictive power. Rather than training a single foundation model for end-to-end survival prediction, the approach leverages a MOE-guided ensemble to manage model callings as tools automatically. By leveraging the strengths of existing models and guiding them through a MOE framework, the aim is to achieve better performance and more accurate predictions in complex real-world tasks. Experiments and analysis on the TCGA-LUAD dataset show improved performance over the individual modal and vanilla ensemble models.

Visual Question Answering (VQA) is an increasingly important multi-modal task where models must answer textual questions based on visual image inputs. Numerous VQA datasets have been proposed to train and evaluate models. However, existing benchmarks exhibit a unilateral focus on textual distribution shifts rather than joint shifts across modalities. This is suboptimal for properly assessing model robustness and generalization. To address this gap, a novel multi-modal VQA benchmark dataset is introduced for the first time. This dataset combines both visual and textual distribution shifts across training and test sets. Using this challenging benchmark exposes vulnerabilities in existing models relying on spurious correlations and overfitting to dataset biases. The novel dataset advances the field by enabling more robust model training and rigorous evaluation of multi-modal distribution shift generalization. In addition, a new few-shot multi-modal prompt fusion model is proposed to better adapt models for downstream VQA tasks. The model incorporates a prompt encoder module and dual-path design to align and fuse image and text prompts. This represents a novel prompt learning approach tailored for multi-modal learning across vision and language. Together, the introduced benchmark dataset and prompt fusion model address key limitations around evaluating and improving VQA model robustness. The work expands the methodology for training models resilient to multi-modal distribution shifts.

Social media platforms provide a rich environment for analyzing user behavior. Recently, deep learning-based methods have been a mainstream approach for social media analysis models involving complex patterns. However, these methods are susceptible to biases in the training data, such as participation inequality. Basically, a mere 1% of users generate the majority of the content on social networking sites, while the remaining users, though engaged to varying degrees, tend to be less active in content creation and largely silent. These silent users consume and listen to information that is propagated on the platform.However, their voice, attitude, and interests are not reflected in the online content, making the decision of the current methods predisposed towards the opinion of the active users. So models can mistake the loudest users for the majority. To make the silent majority heard is to reveal the true landscape of the platform. In this dissertation, to compensate for this bias in the data, which is related to user-level data scarcity, I introduce three pieces of research work. Two of these proposed solutions deal with the data on hand while the other tries to augment the current data. Specifically, the first proposed approach modifies the weight of users' activity/interaction in the input space, while the second approach involves re-weighting the loss based on the users' activity levels during the downstream task training. Lastly, the third approach uses large language models (LLMs) and learns the user's writing behavior to expand the current data. In other words, by utilizing LLMs as a sophisticated knowledge base, this method aims to augment the silent user's data.

Machine learning models are increasingly being deployed in real-world applications where their predictions are used to make critical decisions in a variety of domains. The proliferation of such models has led to a burgeoning need to ensure the reliability and safety of these models, given the potential negative consequences of model vulnerabilities. The complexity of machine learning models, along with the extensive data sets they analyze, can result in unpredictable and unintended outcomes. Model vulnerabilities may manifest due to errors in data input, algorithm design, or model deployment, which can have significant implications for both individuals and society. To prevent such negative outcomes, it is imperative to identify model vulnerabilities at an early stage in the development process. This will aid in guaranteeing the integrity, dependability, and safety of the models, thus mitigating potential risks and enabling the full potential of these technologies to be realized. However, enumerating vulnerabilities can be challenging due to the complexity of the real-world environment. Visual analytics, situated at the intersection of human-computer interaction, computer graphics, and artificial intelligence, offers a promising approach for achieving high interpretability of complex black-box models, thus reducing the cost of obtaining insights into potential vulnerabilities of models. This research is devoted to designing novel visual analytics methods to support the identification and analysis of model vulnerabilities. Specifically, generalizable visual analytics frameworks are instantiated to explore vulnerabilities in machine learning models concerning security (adversarial attacks and data perturbation) and fairness (algorithmic bias). In the end, a visual analytics approach is proposed to enable domain experts to explain and diagnose the model improvement of addressing identified vulnerabilities of machine learning models in a human-in-the-loop fashion. The proposed methods hold the potential to enhance the security and fairness of machine learning models deployed in critical real-world applications.

Artificial Intelligence (AI) systems have achieved outstanding performance and have been found to be better than humans at various tasks, such as sentiment analysis, and face recognition. However, the majority of these state-of-the-art AI systems use complex Deep Learning (DL) methods which present challenges for human experts to design and evaluate such models with respect to privacy, fairness, and robustness. Recent examination of DL models reveals that representations may include information that could lead to privacy violations, unfairness, and robustness issues. This results in AI systems that are potentially untrustworthy from a socio-technical standpoint. Trustworthiness in AI is defined by a set of model properties such as non-discriminatory bias, protection of users’ sensitive attributes, and lawful decision-making. The characteristics of trustworthy AI can be grouped into three categories: Reliability, Resiliency, and Responsibility. Past research has shown that the successful integration of an AI model depends on its trustworthiness. Thus it is crucial for organizations and researchers to build trustworthy AI systems to facilitate the seamless integration and adoption of intelligent technologies. The main issue with existing AI systems is that they are primarily trained to improve technical measures such as accuracy on a specific task but are not considerate of socio-technical measures. The aim of this dissertation is to propose methods for improving the trustworthiness of AI systems through representation learning. DL models’ representations contain information about a given input and can be used for tasks such as detecting fake news on social media or predicting the sentiment of a review. The findings of this dissertation significantly expand the scope of trustworthy AI research and establish a new paradigm for modifying data representations to balance between properties of trustworthy AI. Specifically, this research investigates multiple techniques such as reinforcement learning for understanding trustworthiness in users’ privacy, fairness, and robustness in classification tasks like cyberbullying detection and fake news detection. Since most social measures in trustworthy AI cannot be used to fine-tune or train an AI model directly, the main contribution of this dissertation lies in using reinforcement learning to alter an AI system’s behavior based on non-differentiable social measures.

Graph-structured data, ranging from social networks to financial transaction networks, from citation networks to gene regulatory networks, have been widely used for modeling a myriad of real-world systems. As a prevailing model architecture to model graph-structured data, graph neural networks (GNNs) has drawn much attention in both academic and industrial communities in the past decades. Despite their success in different graph learning tasks, existing methods usually rely on learning from ``big'' data, requiring a large amount of labeled data for model training. However, it is common that real-world graphs are associated with ``small'' labeled data as data annotation and labeling on graphs is always time and resource-consuming. Therefore, it is imperative to investigate graph machine learning (Graph ML) with low-cost human supervision for low-resource settings where limited or even no labeled data is available. This dissertation investigates a new research field -- Data-Efficient Graph Learning, which aims to push forward the performance boundary of graph machine learning (Graph ML) models with different kinds of low-cost supervision signals. To achieve this goal, a series of studies are conducted for solving different data-efficient graph learning problems, including graph few-shot learning, graph weakly-supervised learning, and graph self-supervised learning.

As threats emerge and change, the life of a police officer continues to intensify. To better support police training curriculums and police cadets through this critical career juncture, this thesis proposes a state-of-the-art framework for stress detection using real-world data and deep neural networks. As an integral step of a larger study, this thesis investigates data processing techniques to handle the ambiguity of data collected in naturalistic contexts and leverages data structuring approaches to train deep neural networks. The analysis used data collected from 37 police training cadetsin five different training cohorts at the Phoenix Police Regional Training Academy. The data was collected at different intervals during the cadets’ rigorous six-month training course. In total, data were collected over 11 months from all the cohorts combined. All cadets were equipped with a Fitbit wearable device with a custom-built application to collect biometric data, including heart rate and self-reported stress levels. Throughout the data collection period, the cadets were asked to wear the Fitbit device and respond to stress level prompts to capture real-time responses. To manage this naturalistic data, this thesis leveraged heart rate filtering algorithms, including Hampel, Median, Savitzky-Golay, and Wiener, to remove potentially noisy data. After data processing and noise removal, the heart rate data and corresponding stress level labels are processed into two different dataset sizes. The data is then fed into a Deep ECGNet (created by Prajod et al.), a simple Feed Forward network (created by Sim et al.), and a Multilayer Perceptron (MLP) network for binary classification. Experimental results show that the Feed Forward network achieves the highest accuracy (90.66%) for data from a single cohort, while the MLP model performs best on data across cohorts, achieving an 85.92% accuracy. These findings suggest that stress detection is feasible on a variate set of real-world data using deepneural networks.