AI BriefingLG AI Research 256
Key point
LG AI Research introduced key research on self-supervised learning presented at ICML 2022.
Details
LG AI Research participated in ICML 2022, a world-renowned AI academic conference, to share the latest research trends. This blog series covers notable papers based on the two core directions of self-supervised learning: 'Self-prediction' and 'Contrastive learning'.
In particular, the ADIOS methodology introduced proposes a new framework for effectively learning the semantic information of images. Existing Masked Image Modeling approaches had limitations in learning the core meaning of images, as the process of randomly masking and restoring pixels relied only on local correlations with surrounding pixels.
ADIOS has a structure in which the Inference model and Occlusion model learn while competing against each other to solve this problem.
- Occlusion model (M): Generates a mask the size of the original image, and learns to maximize the distance between the masked version and the original representation.
- Inference model (I): Takes the masked image and the original image as input to compute each representation, and learns to minimize the distance between the two representations.
Through this adversarial learning approach, the model can more effectively learn parts containing important semantic information of the image, rather than simply replicating surrounding pixels.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.