NUMERICAL METHODS AND SOFTWARE COMPLEX OF ADAPTIVE ADJUSTMENT OF IMAGE-LANGUAGE MODELS IN UZBEK LANGUAGE
This article proposes numerical methods and a software architecture for adaptive optimization of Uzbek vision-language models under limited computational resources. The mathematical formulation combines contrastive alignment, image-conditioned text likelihood, and regularization. Low-rank parameterization, AdamW iterations, gradient clipping, and validation-based step-size control are described. The responsibilities of data processing, numerical computation, state persistence, retrieval, and evaluation modules are specified. Analytical memory estimates are provided for parameter states and the similarity matrix. Two equivalent implementations of a low-rank operator were compared in a numerical check, with a maximum absolute difference below 1.34×10⁻¹⁵. This result demonstrates algebraic consistency of the operator rather than model accuracy or runtime gains. A reproducible experimental protocol is defined for evaluating performance and resource consumption on real data.
Authors
- Muhammadbobur Mirzaaxmedov
- Otakishiyeva Gulshanoy Abdulaziz qizi
Publication Details
- Journal
- International journal of technology and academic research
- Published
- 2026-11-12
- Primary Topic
- Multimodal Machine Learning Applications
- Type
- article
- Field-Weighted Citation Impact
- 0.00