1,720,958 research outputs found
Recommended from our members
From Understanding to Improving Artificial Intelligence: New Frontiers in Machine Learning Explanations
No full textAs machine learning systems increasingly shape outcomes in high-stakes domains, the need to understand, trust, and effectively guide their decision-making grows urgent. This dissertation advances the field of machine learning explainability, offering a cohesive framework for enabling AI systems whose underlying reasoning is transparent, resilient, and actionable. By examining three critical frontiers—explainability amidst adversarial robustness, scalable rationale generation for large language models (LLMs), and decoding LLM behavior under iterative prompting—this work illuminates how explanations can inform, protect, and empower stakeholders.
The first part reveals how adversarial training, while bolstering model security, can inadvertently undermine the provision of meaningful, low-cost algorithmic recourse. This tension exposes trade-offs between securing decision boundaries and preserving explanations that help individuals improve their predicted outcomes. The second part introduces a novel approach to scaling explanations without human annotation, integrating post hoc attributions from smaller, more interpretable proxy models directly into LLM prompting. This not only reduces the need for manual rationales but also demonstrates that automatically generated explanations can actively guide complex models toward more coherent and well-founded reasoning.
The final part focuses on decoding LLM behavior through iterative prompting. While one might expect repeated user-model interactions to improve understanding and truthfulness, naïve iterative prompting can paradoxically degrade factual alignment and confidence calibration. By carefully analyzing how LLMs respond to iterative queries, the dissertation uncovers new insights into model tendencies, including over-apologizing and sycophantic patterns, and develops strategies to mitigate these issues. This examination shows that how we interact with models—how we request, refine, and interpret explanations—fundamentally shapes model reliability and clarity.
Collectively, these contributions emphasize that robust, scalable, and iteratively refined explanations are both feasible and vital. By reconciling adversarial defenses with user-friendly recourse, automating rationales for complex models, and decoding LLM behaviors through iterative engagement, the dissertation provides a principled path toward AI systems whose inner workings can be understood, trusted, and responsibly guided by human stakeholders.Engineering and Applied Sciences - Computer Scienc
Recommended from our members
Epistemic Limits of Trustworthy Machine Learning
Get PDFTheoretical understanding of a system’s limits has long driven technological breakthroughs. Carnot delineated the fundamental limits of heat engine efficiency, paving the way for the design of modern state-of-the-art engines. More than a century later, Claude Shannon unraveled the fundamental limit of communication, known as channel capacity. This insight revolutionized communication systems, enabling continual improvements that ultimately led to wireless communication as we know it today.
This thesis discusses the epistemic limits of machine learning (ML) and leverages them to improve the trustworthiness of ML systems. ML models have an epistemic limit when proving one of their properties is impossible. Epistemic refers to the impossibility of providing theoretical guarantees (knowledge) about a model's property. Epistemic limits are information-theoretic converse results on the hypothesis test that checks a model's property.
First, we prove a limit on how much information personalized models can use while ensuring reliable test for performance gains across all users -- epistemic limits of personalization. We leverage this limit to develop a tool to help with feature selection. Second, we show a limit for reliably testing if model performance is equitable across multiple demographic groups --epistemic limit of fairness testing. We exploit this limit to design a metric for efficient algorithmic bias detection.
Third, we prove a limit for testing if one model outperforms another on average -- epistemic limit of model selection.
We use this result to delineate the set of indistinguishably good models --Rashomon set. Finally, we argue that the epistemic limits in model selection imply that explaining the predictions of ML models is necessary. Then, we develop efficient methods for explaining the content produced by large language models.Engineering and Applied Sciences - Applied Mat
Recommended from our members
Deriving Indistinguishability from Unpredictability: Tools and Applications in Pseudorandomness
No full textProving that a distribution P is “close to uniform” is an integral part of many problems in pseudorandomness, and is often defined either in terms of indistinguishability—no algorithm (possibly required to be efficient) should be able to distinguish between P and the uniform distribution, or unpredictability—the distribution P should have high entropy, or be unpredictable by any (efficient) algorithm. In most cases, the application will require the former type of guarantee, although the latter can sometimes be easier to reason about. In this thesis, we develop tools to relate these notions, and apply information-theoretic reasoning to problems in complexity theory and cryptography:
• We extend the definition of randomness extractors to allow the error to be measured in terms of an arbitrary distance measure, and extend the connection between extractors and averaging samplers (Zuckerman, Rand. Struct. Alg.‘97) to an arbitrary family ℱ of test functions and the integral probability metric defined by ℱ. Using this connection, we show that extractors for the Kullback–Leibler (KL) divergence are subgaussian samplers as defined by Błasiok (SODA‘18). By showing that KL extractors exist with essentially the same parameters as standard extractors (explicitly and non-explicitly), we construct the first explicit subgaussian samplers matching the best known constructions of averaging samplers for [0, 1]-bounded functions in the parameter regime where the approximation error ε and failure probability δ are subconstant.
• We introduce hardness in relative entropy, a new notion of hardness for search problems which on the one hand is satisfied by all one-way functions and on the other hand implies both next-block pseudoentropy and inaccessible entropy, two forms of computational entropy used in recent constructions of pseudorandom generators and statistically hiding commitment schemes, respectively, thereby shedding light on the apparent “duality” between them.
• We show that the moment generating function of the KL divergence between the empirical distribution of n independent samples from a distribution P over a finite alphabet of size k (i.e. a multinomial distribution) and P itself is no more than that of a gamma distribution with shape k − 1 and rate n. The resulting exponential concentration inequality becomes meaningful (less than 1) when the divergence ε is larger than (k − 1)/n, whereas the standard method of types bound requires ε > 1/n · log(n+k−1 choose k−1) ≥ (k − 1)/n · log(1 + n/(k − 1)), thus saving a factor of order log(n/k) in the standard regime of parameters where n ≫ k.
• We systematically study the relationship between f-divergences and integral probability metrics (IPMs) from the perspective of convex duality. Starting from a tight variational representation of the f-divergence, we derive a generalization of the moment generating function, which we show exactly characterizes the best lower bound of the f-divergence as a function of a given IPM. Using this characterization, we obtain new bounds while also recovering in a unified manner well-known results, such as Hoeffding’s lemma, Pinsker’s inequality and its extension to subgaussian functions, and the Hammersley–Chapman–Robbins bound. The variational representation also allows us to prove new results on topological properties of the divergence which may be of independent interest
Recommended from our members
Towards Practical Applications of Machine Learning in Healthcare with Federated Learning
No full textFederated Learning (FL) has emerged as a significant tool in healthcare machine learning, enabling institutions to collaboratively train models while maintaining data privacy. This dissertation describes the implementation of a real-world healthcare FL project and addresses the challenge of domain shift for more effective model deployment.
We begin by detailing a practical application of FL during the SARS-COV-2 pandemic. Twenty institutions collaborated on a healthcare FL study to develop the "EXAM" (EMR CXR AI Model), which predicts future oxygen requirements for symptomatic patients using vital signs, laboratory data, and chest x-rays. EXAM achieved an Area Under the Curve (AUC) of over 0.92, marking a 16\% improvement and a 38\% increase in generalizability over local models. This project demonstrated FL's ability to enable rapid scientific collaboration without data exchange, producing a model that generalized across heterogeneous, unharmonized datasets and provided the healthcare community with a validated tool to combat COVID-19.
Next, we address a specialized non-iid FL challenge termed \emph{Domain-mixed FL}, where each client's data is assumed to be a mixture of several predefined domains. We propose a novel method, FedDAR, which learns a shared domain representation and personalized prediction models in a decoupled manner. Theoretical proofs show that FedDAR achieves linear convergence in simplified settings, and extensive empirical studies on both synthetic and real-world datasets demonstrate its superiority over existing FL methods.
Finally, we explore the multi-dimensional domain shift problem prevalent in healthcare ML applications. We introduce a novel strategy using an ensemble of mixtures of experts (EMoE), each expert tailored to adapt to shifts along different dimensions. This approach is designed to be versatile and robust, suitable for both centralized and federated learning settings. Rigorous testing on various real-world datasets has shown that our method outperforms contemporary domain generalization and personalized federated learning approaches, effectively managing the complexities of multi-dimensional domain shifts
Going Beyond Counting First Authors in Author Co-citation Analysis
Get PDFThe present study examines one of the fundamental aspects of author co-citation analysis (ACA) - the way co-citation
counts are defined. Co-citation counting provides the data on which all subsequent statistical analyses and mappings
are based, and we compare ACA results based on two different types of co-citation counting - the traditional type that
only counts the first one among a cited work's authors on the one hand and a non-traditional type that takes into
account the first 5 authors of a cited work on the other hand. Results indicate that the picture produced through this non-traditional author co-citation counting contains more coherent author groups and is therefore considerably clearer. However, this picture represents fewer specialties in the research field being studied than that produced through the traditional first-author co-citation counting when the same number of top-ranked authors is selected and analyzed. Reasons for these effects are discussed
Variations on the Author
Get PDF“Variations on the Author” discusses two of Eduardo Coutinho’s recent films (Um Dia na Vida, from 2010, and Últimas Conversas, posthumously released in 2015) and their contribution to the general question of documentary authorship. The director’s filmography is characterized by a consistent yet self-effacing form of authorial self-inscription: Coutinho often features as an interviewer that rather than express opinions propels discourses; an interviewer that is good at listening. This mode of self-inscription characterizes him as an author who is not expressive but who is nonetheless markedly present on the screen. In Um Dia na Vida, however, Coutinho is completely absent form the image, while Últimas Conversas, on the contrary, includes a confessional prologue that moves the director from the margins to the center of his films. This article examines the ways in which these works stand out in the filmography of a director who offers new insights into the notion of cinematic authorship
Appropriate Similarity Measures for Author Cocitation Analysis
Get PDFWe provide a number of new insights into the methodological discussion about author cocitation analysis. We first argue that the use of the Pearson correlation for measuring the similarity between authors’ cocitation profiles is not very satisfactory. We then discuss what kind of similarity measures may be used as an alternative to the Pearson correlation. We consider three similarity measures in particular. One is the well-known cosine. The other two similarity measures have not been used before in the bibliometric literature. Finally, we show by means of an example that our findings have a high practical relevance.information science;Pearson correlation;cosine;similarity measure;author cocitation analysis
Dispelling the Myths Behind First-author Citation Counts
Get PDFWe conducted a full-scale evaluative citation analysis study of scholars in the XML research field to explore just how different from each other author rankings resulting from different citation counting methods actually are, and to demonstrate the capability of emerging data and tools on the Web in supporting more realistic citation counting methods. Our results contest some common arguments for the continued
use of first-author citation counts in the evaluation of scholars, such as high correlations between author rankings by first-author citation counts and other citation
counting methods, and high costs of using more realistic citation counting methods that are not well-supported by the ISI databases. It is argued that increasingly available digital full text research papers make it possible for citation analysis studies to go beyond what the ISI databases have directly supported and to employ more
sophisticated methods
- …
