MORE
#blip #interview #salesforce
Paper Review Video: https://youtu.be/X2k7n4FuI7c
Sponsor: Assembly AI
https://www.assemblyai.com/?utm_sourc...
This is an interview with Junnan Li and Dongxu Li, authors of BLIP and members of Salesforce research.
Cross-modal pre-training has been all the rage lately in deep learning, especially training vision and language models together. However, there are a number of issues, such as low quality datasets that limit the performance of any model trained on it, and also the fact that pure contrastive pre-training cannot be easily fine-tuned for most downstream tasks. BLIP unifies different tasks and objectives in a single pre-training run and achieves a much more versatile model, which the paper immediately uses to create, filter, clean and thus bootstrap its own dataset to improve performance even more!
OUTLINE:
0:00 - Intro
0:40 - Sponsor: Assembly AI
1:30 - Start of Interview
2:30 - What's the pitch?
4:40 - How did data bootstrapping come into the project?
7:10 - How big of a problem is data quality?
11:10 - Are the captioning & filtering models biased towards COCO data?
14:40 - Could the data bootstrapping be done multiple times?
16:20 - What was the evolution of the BLIP architecture?
21:15 - Are there additional benefits to adding language modelling?
23:50 - Can we imagine a modular future for pre-training?
29:45 - Diving into the experimental results
42:40 - What did and did not work out during the research?
45:00 - How is research life at Salesforce?
46:45 - Where do we go from here?
Paper: https://arxiv.org/abs/2201.12086
Code: https://github.com/salesforce/BLIP
Demo: https://huggingface.co/spaces/Salesfo...
Abstract:
Vision-Language Pre-training (VLP) has advanced the performance for many vision-language tasks. However, most existing pre-trained models only excel in either understanding-based tasks or generation-based tasks. Furthermore, performance improvement has been largely achieved by scaling up the dataset with noisy image-text pairs collected from the web, which is a suboptimal source of supervision. In this paper, we propose BLIP, a new VLP framework which transfers flexibly to both vision-language understanding and generation tasks. BLIP effectively utilizes the noisy web data by bootstrapping the captions, where a captioner generates synthetic captions and a filter removes the noisy ones. We achieve state-of-the-art results on a wide range of vision-language tasks, such as image-text retrieval (+2.7% in average recall@1), image captioning (+2.8% in CIDEr), and VQA (+1.6% in VQA score). BLIP also demonstrates strong generalization ability when directly transferred to video-language tasks in a zero-shot manner. Code, models, and datasets are released at this https URL.
Authors: Junnan Li, Dongxu Li, Caiming Xiong, Steven Hoi
