## Introduction and Motivation
Visual Question Answering (VQA) is a task where we expect the AI to answer a question about a given image. VQA has been an active area of research for the past 4-5 years, with most datasets using natural images found online. Two examples of such datasets are: [VQAv2](https://visualqa.org/challenge.html), [GQA](https://cs.stanford.edu/people/dorarad/gqa/about.html). VQA is a particularly interesting multi-modal machine learning challenge because it has several interesting applications across several domains including healthcare chatbots, interactive-agents, etc. **However, most VQA challenges or datasets deal with English-only captions and questions.** 

In addition, even recent **approaches that have been proposed for VQA generally are obscure** due to the fact that CNN-based object detectors are relatively difficult to use and more complex for feature extraction. For example, a FasterRCNN approach uses the following steps:
- a FPN (Feature Pyramid Net) over a ResNet backbone, and
- then a RPN (Regision Proposal Network) layer detects proposals in those features, and
- then the ROI (Region of Interest) heads get the box proposals in the original image, and
- the the boxes are selected using a NMS (Non-max suppression),
- and then the features for selected boxes are used as visual features.

A major **advantage that comes from using transformers is their simplicity and their accessibility** - thanks to HuggingFace team, ViT and Transformers authors. For ViT models, for example, all one needs to do is pass the normalized images to the transformer.

While building a low-resource non-English VQA approach has several benefits of its own, a multilingual VQA task is interesting because it will help create a generic approach/model that works decently well across several languages.

**With the aim of democratizing such an challenging yet interesting task, in this project, we focus on Mutilingual Visual Question Answering (MVQA)**. Our intention here is to provide a Proof-of-Concept with our simple CLIP-Vision-BERT baseline which leverages a multilingual checkpoint with pre-trained image encoders. Our model currently supports for four languages - **English, French, German and Spanish**.

We follow the two-staged training approach, our pre-training task being text-only Masked Language Modeling (MLM). Our pre-training dataset comes from Conceptual-12M dataset where we use mBART-50 for translation. Our fine-tuning dataset is taken from the VQAv2 dataset and its translation is done using MarianMT models.

Our checkpoints achieve a **validation accuracy of 0.69 on our MLM** task, while our fine-tuned model is able to achieve a **validation accuracy of 0.49 on our multilingual VQAv2 validation set**. With better captions, hyperparameter-tuning, and further training, we expect to see higher performance.

### Novel Contributions
Our novel contributions include:
- A [multilingual variant of the Conceptual-12M dataset](https://huggingface.co/datasets/flax-community/conceptual-12m-mbart-50-multilingual) containing 2.5M image-text pairs each in four languages - English, French, German and Spanish, translated using mBART-50 model.
- [Multilingual variants of the VQAv2 train and validation sets](https://huggingface.co/datasets/flax-community/multilingual-vqa) containing four times the original data in English, French, German and Spanish, translated using Marian models.
- [A fusion of CLIP Vision Transformer and BERT model](https://github.com/gchhablani/multilingual-vqa/tree/main/models/flax_clip_vision_bert) where BERT embeddings are concatenated with visual embeddings at the very beginning and passed through BERT self-attention layers. This is based on the [VisualBERT](https://arxiv.org/abs/1908.03557) model.
- A [pre-trained checkpooint](https://huggingface.co/flax-community/clip-vision-bert-cc12m-70k) on our multilingual with **0.69** validation accuracy.
- A [fine-tuned checkpoint](https://huggingface.co/flax-community/clip-vision-bert-vqa-ft-6k) on our multilingual variant of the VQAv2 dataset with **0.49** validation accuracy.