JeanIrad/dl_project23

Investigating how to dynamically steer Vision-Language Models (VLMs) using language prompts to overcome geographical biases and improve African cultural recognition. # 🌍 Dynamic Perceptual Steering **Orchestrating Shape and Texture Biases in Vision Language Models for African Cultural Competency** > This repository contains the code, data curation scripts, and evaluation framework for the project: **Dynamic Perceptual Steering**. We investigate how to use language prompts to dynamically steer Vision-Language Models (VLMs) to recognize African cultural cues, overcoming inherent Western-centric geographical biases. ## 📑 Table of Contents - Abstract - Motivation & Objectives - Background & Related Work - Methodology - Datasets & Evaluation - Baselines & Experiments - Expected Results & Analysis - Project Roadmap - Team Contributions - References --- ## 📖 Abstract Vision-Language Models (VLMs) are increasingly used to describe and categorize the visual world, yet they suffer from a significant geographical bias. Primarily trained on Western data, these models often "perceive" African objects through a Western lens—either misidentifying them or ignoring their unique cultural details. This project investigates whether we can "talk" these models into seeing differently by using specific language prompts to steer their focus. We propose a **"Dynamic Steering"** framework that allows a model to first use its "shape bias" to identify what an object is (e.g., a bowl) and then switch to its "texture bias" to recognize its cultural origin (e.g., a specific African artisan style). --- ## 🎯 Motivation & Objectives ### The Problem: Perceptual Erasure As AI tools are deployed globally, they risk **"perceptual erasure"**—the failure to recognize the cultural and material diversity of the Global South. Current models are often "stubborn," favoring universal shapes over local textures, leading to a loss of important cultural context in applications like digital heritage, education, and e-commerce. ### Broad Objectives: 1. **Measure Stubbornness:** Quantify how resistant current VLMs are when faced with African-specific textures an …