absorb.md

Ilyas Khan

Chronological feed of everything captured from Ilyas Khan.

Vision-Language Models Exhibit Geographical Bias in Indian Face Recognition

Vision-Language Models (VLMs) demonstrate significant geographical bias within their representations of Indian demographics, stemming from the oversimplification of India's diverse population in existing fairness datasets. The IndicFairFace dataset, a novel and balanced compilation of 14,400 images, was developed to address this limitation by uniformly representing India's diverse states and genders. This dataset enables quantification and mitigation of intra-national geographical bias in VLMs with minimal impact on model performance.

Pictorial Mathematics Lowers Barrier to Quantum Education

Quantum Picturalism, a novel visual mathematical language, has demonstrated significant potential as an educational methodology. A pilot study showed high school students, without advanced mathematical prerequisites, achieved an 82% pass rate on university-level quantum exams. This approach could broaden participation in quantum computing by lowering traditional cognitive and demographic barriers.

Diagrammatic Quantum Mechanics Enables High School-Level QIST Education Without Advanced Math Prerequisites

Quantum Picturalism (QPic) is a fully diagrammatic formalism covering all of qubit quantum mechanics — including entanglement, measurement, and mixed states — that eliminates the need for matrices, vectors, tensors, complex numbers, and trigonometry. A controlled study with 54 high school students (ages 16–18) demonstrated that 16 hours of QPic training over eight weeks produced strong outcomes: 82% passed a final exam and 48% earned a distinction. This suggests that Quantum Information Science and Technology (QIST), traditionally restricted to university-level education, can be rigorously introduced at the secondary level through formalism redesign rather than content simplification.

Compositional Interpretability for Explainable AI via Category Theory

This paper introduces a novel approach to Explainable AI (XAI) by defining AI models and their interpretability using category theory. The authors propose "compositional models" which use formal string diagrams to represent both the abstract structure and concrete implementation of various AI models. This framework facilitates diagrammatic explanations and allows for the identification of compositionally-interpretable (CI) models, offering benefits for understanding and inferring behavior from complex AI systems.

Uniqueness of Asymptotically Conical Gradient Shrinking Solitons in G₂-Laplacian Flow

This paper presents a proof of the uniqueness of asymptotically conical (AC) gradient shrinking solitons for the Laplacian flow of closed G₂-structures. The authors extend existing arguments for Ricci solitons to the context of G₂-Laplacian flow. The research also investigates the inheritance of symmetries from the asymptotic cone to the G₂-structure of an AC shrinker end.