Deep Learning

The system utilizes state-of-the-art natural language processing techniques like BERT (Bidirectional Encoder Representations from Transformers) to understand the user’s question and find the best matching question-answer pair. Fine-tuning BERT on domain-specific data allows the system to become adept at understanding queries in a particular domain.

Used Pretrained CLIP image encoder and GPT-2 Language Model to build a multilingual image captioning system on English, Arabic, French, and Deutsch. The system was trained on 1.5M images and 4.5M captions from the Multi30K dataset, with 4 prefix adapters each on a specific language. The system was able to generate captions in the four languages with high accuracy.

This project fine-tuned compact monolingual code language models to new programming languages using efficient techniques, with the goal of expanding capabilities and accessibility of code intelligence tools. The work shares insights from optimizing training efficiency and makes models/datasets publicly available to further AI democratization.

This study aims to bridge this research gap through the development of a robust methodology that leverages remote sensing imagery and machine learning classification. We present a comprehensive model, which we employed to derive estimates of cultivated agricultural land in Gezira state, spanning the timeframe from 2019 to 2023.