MoDE: CLIP Data Experts via Clustering
This addresses noisy web-crawled data for CLIP training, offering a scalable and efficient solution for improving zero-shot performance in vision-language models.
The paper tackles noisy supervision in CLIP training by introducing Mixture of Data Experts (MoDE), which trains multiple CLIP models on clustered data to reduce sensitivity to false negatives, resulting in four experts on ViT-B/16 outperforming larger models like ViT-L/14 in zero-shot image classification with less than 35% training cost.
The success of contrastive language-image pretraining (CLIP) relies on the supervision from the pairing between images and captions, which tends to be noisy in web-crawled data. We present Mixture of Data Experts (MoDE) and learn a system of CLIP data experts via clustering. Each data expert is trained on one data cluster, being less sensitive to false negative noises in other clusters. At inference time, we ensemble their outputs by applying weights determined through the correlation between task metadata and cluster conditions. To estimate the correlation precisely, the samples in one cluster should be semantically similar, but the number of data experts should still be reasonable for training and inference. As such, we consider the ontology in human language and propose to use fine-grained cluster centers to represent each data expert at a coarse-grained level. Experimental studies show that four CLIP data experts on ViT-B/16 outperform the ViT-L/14 by OpenAI CLIP and OpenCLIP on zero-shot image classification but with less ($<$35\%) training cost. Meanwhile, MoDE can train all data expert asynchronously and can flexibly include new data experts. The code is available at https://github.com/facebookresearch/MetaCLIP/tree/main/mode.