MoCo and Contrastive Learning

less than 1 minute read

Basic ideas in Contrastive Learning and Kaiming’s improvment in MoCo. Alt text

0 Contrastive Learning

The overview position of CL Alt text A slide from Yang LeCun Alt text

To do CL is supposed to be easiler than SL. Human is easy to distinguish objects rather than identify objects. The idea is similar to training embedding with triplets. Alt text So with postive and negative samples, you are trained to distinguish them too. Alt text

1 Memory Bank

A earlier paper proposed memory bank idea. Alt text

Alt text

2 Momentum Contrastive

Kaiming combined the end-to-end and memory bank and proposed momentum contrastive. Alt text

The loss is defined as InfoNCE Alt text Alt text Try to meet both alignment and uniformaity goal. Alt text

Comparing to Memory Bank, it’s similar to momentum in BP learning. Alt text

Tags:

Categories:

Updated: