Fairness and Explainability for Enabling Trust in AI Systems

Human-computer interaction series, Год журнала: 2024, Номер unknown, С. 85 - 110

Опубликована: Янв. 1, 2024

Язык: Английский

---

Evolving Interpretable Visual Classifiers with Large Language Models

Lecture notes in computer science, Год журнала: 2024, Номер unknown, С. 183 - 201

Опубликована: Окт. 30, 2024

Язык: Английский

---

FedMUA: Exploring the Vulnerabilities of Federated Learning to Malicious Unlearning Attacks

IEEE Transactions on Information Forensics and Security, Год журнала: 2025, Номер 20, С. 1665 - 1678

Опубликована: Янв. 1, 2025

Язык: Английский

---

Conversational Explanations: Discussing Explainable AI with Non-AI Experts

Опубликована: Март 19, 2025

Язык: Английский

---

Class-wise federated unlearning: Harnessing active forgetting with teacher-student memory generation

Knowledge-Based Systems, Год журнала: 2025, Номер unknown, С. 113353 - 113353

Опубликована: Март 1, 2025

Язык: Английский

---

Responsible AI

Human-computer interaction series, Год журнала: 2025, Номер unknown, С. 619 - 644

Опубликована: Янв. 1, 2025

---

Interpreting Deep Learning Models in Natural Language Processing: A Review

arXiv (Cornell University), Год журнала: 2021, Номер unknown

Опубликована: Янв. 1, 2021

Neural network models have achieved state-of-the-art performances in a wide range of natural language processing (NLP) tasks. However, long-standing criticism against neural is the lack interpretability, which not only reduces reliability NLP systems but also limits scope their applications areas where interpretability essential (e.g., health care applications). In response, increasing interest interpreting has spurred diverse array interpretation methods over recent years. this survey, we provide comprehensive review various for NLP. We first stretch out high-level taxonomy NLP, i.e., training-based approaches, test-based and hybrid approaches. Next, describe sub-categories each category detail, e.g., influence-function based methods, KNN-based attention-based models, saliency-based perturbation-based etc. point deficiencies current suggest some avenues future research.

Язык: Английский

---

HILDIF: Interactive Debugging of NLI Models Using Influence Functions

Опубликована: Янв. 1, 2021

Biases and artifacts in training data can cause unwelcome behavior text classifiers (such as shallow pattern matching), leading to lack of generalizability. One solution this problem is include users the loop leverage their feedback improve models. We propose a novel explanatory debugging pipeline called HILDIF, enabling humans deep using influence functions an explanation method. experiment on Natural Language Inference (NLI) task, showing that HILDIF effectively alleviate artifact problems fine-tuned BERT models result increased model

Язык: Английский

---

Evaluating Data Attribution for Text-to-Image Models

2021 IEEE/CVF International Conference on Computer Vision (ICCV), Год журнала: 2023, Номер unknown, С. 7158 - 7169

Опубликована: Окт. 1, 2023

While large text-to-image models are able to synthesize "novel" images, these images necessarily a reflection of the training data. The problem data attribution in such – which set most responsible for appearance given generated image is difficult yet important one. As an initial step toward this problem, we evaluate through "customization" methods, tune existing large-scale model exemplar object or style. Our key insight that allow us efficiently create synthetic computationally influenced by construction. With our new dataset exemplar-influenced various algorithms and different possible feature spaces. Furthermore, on dataset, can standard models, as DINO, CLIP, ViT, problem. Even though procedure tuned towards small sets, show generalization larger sets. Finally, taking into account inherent uncertainty assign soft scores over images.

Язык: Английский

---

Neural networks memorise personal information from one sample

Scientific Reports, Год журнала: 2023, Номер 13(1)

Опубликована: Дек. 4, 2023

Deep neural networks (DNNs) have achieved high accuracy in diagnosing multiple diseases/conditions at a large scale. However, number of concerns been raised about safeguarding data privacy and algorithmic bias the network models. We demonstrate that unique features (UFs), such as names, IDs, or other patient information can be memorised (and eventually leaked) by even when it occurs on single training sample within dataset. explain this memorisation phenomenon showing is more likely to occur UFs are an instance rare concept. propose methods identify whether given model does not memorise (known) feature. Importantly, our method require access therefore deployed external entity. conclude implications robustness, but also pose risk patients who consent use their for

Язык: Английский

---