New ideas for BTP: Visual Question Answering

IdeaList.md

• 1x1 convolutions have been extensively used to reduce the number of parameters without affecting the results much
• Deep Mutual Learning: Unlike bagging/ boosting, models learn jointly, and help each other to fit well
• Skip connections: Help solving degradation problem without adding parameters.
• Hard Sample Mining

List of some fusion techniques [This could help as we will have to fuse two branches i.e image and text.]

1. https://arxiv.org/pdf/2003.10758.pdf
2. https://arxiv.org/pdf/2103.09422v1.pdf
3. http://www.zhengzhu.net/upload/P6938bc861e8d4583bf47d47d64ed9598.pdf
4. concatenation
5. convolution

Space Efficient tensor operations:

1. https://arxiv.org/pdf/1901.11261.pdf

1. There exist better evaluation metrics than rule based ones (BLEU, CIDER) as @akhileshkb was mentioning yesterday. reference
2. Loss functions similar to classification problem might not be the best idea. Rather exploration of loss function which takes into account the meaning of words might be helpful.

Would create a repo soon, so that all the discussion goes into discussion section of the repo

Knowledge distillation for building small models (small num of parameters) with high performance.

Discussion moved to vqa/wiki