toc: true title: MoCO
tags: [‘temp’]
MoCO
- Momentum Contrast for Unsupervised Visual Representation Learning
- unsupervised visual representation learning
- contrastive learning as dictionary look-up, MoCo builds a dynamic dictionary with a queue and a moving-averaged encoder
- large and consistent dictionary on-the-fly
- ImageNet
- transfer well to downstream tasks.
- PASCAL VOC
- COCO
- visual representation encoder by matching an encoded query
- to a dictionary of encoded keys using a Contrastive Loss
- dictionary is built as a queue, with the current mini-batch enqueued
- oldest mini-batch dequeued
- slowly progressing encoder
- momentum update with the query encoder
