In order to solve the endogenous contradiction between the data scale dividend and the diminishing marginal effect of computing power in large-scale deep learning, this paper proposes a collaborative learning framework for large-scale Dataset distillation and Filtering (DF-CoLearn). By constructing a dynamic feedback closed loop based on bi-level optimization and mutual information maximization, the Pareto optimality between training efficiency and model generalization ability is realized, which provides a new theoretical perspective and technical path for green and efficient AI model training.

Show more

Measuring dietary exposure is the key aspect of nutritional epidemiology in order to find cause and effect relationships between nutrition and long-term illnesses. Nevertheless, self-reported nutrition assessment tools e.g. food frequency questionnaires and dietary recalls provide systematic underreporting as well as random error in nutrition assessment that considerably reduce the regression coefficients of exposure-outcome relationships and even obscure true diet-health effects under measurement errors. The current corrections methods conducted on small reference samples and depending on the assumptions of linearity are capable of treating variations of errors in multimodal data of great dimensions. We are going to present an idea of self-supervised multimodal representation learning, that is, an error-reducting dietary exposure measure, where dietary text logs and wearable sensor data are modeled jointly and learns discriminative features highly correlated with true intake through cross-modal contrastive learning and masked reconstruction, trained over multi-view representations to produce an exposure-corrected dietary energy and nutrient consumption estimate using a unified latent space, and produce an exposure-corrected dietary text log estimate using a unified latent space.

Show more

Image-based artificial intelligence models are widely applied in data science tasks such as image classification, object recognition, and visual content generation. In practice, model outputs are often regarded as reliable once acceptable accuracy levels are achieved on benchmark datasets. However, empirical evidence shows that image-based AI systems frequently exhibit structured and non-random error patterns. In image generation tasks, errors commonly arise from an overreliance on statistical correlations learned from training data, limited semantic grounding, and weak constraints on physical and contextual consistency. These limitations can lead to outputs that appear visually coherent while containing incorrect or non-existent objects, implausible spatial relationships, or violations of basic visual logic. From a data science perspective, such errors are often underexamined because evaluation practices rely heavily on aggregate accuracy metrics and benchmark performance, which tend to obscure localized error concentration and output variability. This study conducts a structured analysis of error patterns and output limitations in image-based AI systems by examining misclassification behavior, generation inconsistencies, and evaluation blind spots observed under realistic data conditions. The findings indicate that understanding AI image generation errors requires focusing on error structure and underlying generation mechanisms rather than relying solely on summary performance measures.

Show more

With the widespread deployment of machine learning models in high-stakes decision-making contexts, their inherent opacity—often termed the "black-box" problem—has raised significant concerns regarding interpretability and reliability. This paper presents a systematic and comprehensive literature review examining the convergence of interpretable machine learning and statistical inference. This paper synthesizes foundational concepts, methodological frameworks, theoretical advancements, and practical applications to elucidate how statistical tools can validate, enhance, and formalize machine learning explanations. This review critically analyzes widely adopted techniques such as SHAP and LIME, and explores their integration with statistical inference tools, including hypothesis testing, confidence intervals, Bayesian methods, and causal inference frameworks. The analysis reveals that integrated approaches significantly improve explanation credibility, regulatory compliance, and decision transparency in critical domains, including healthcare diagnostics, financial risk management, and algorithmic governance. However, persistent challenges remain in theoretical consistency, computational efficiency, evaluation standardization, and human-centered design. This paper concludes by proposing a structured research agenda focusing on unified theoretical frameworks, efficient algorithmic implementations, domain-specific evaluation standards, and interdisciplinary collaboration strategies to advance the responsible development and deployment of explainable AI systems.

Show more