MAE-FS

Masked Autoencoders for Few-Shot Learning (MAE-FS) is a self-supervised, generative technique that reinforces few-shot classification performance for a prototypical backbone model. Given an embedded support set (produced by a frozen backbone), MAE-FS generates new prototypes, through a novel process, all of which are incorporated into class-based centroids. The reinforced centroids are used to classify unlabelled prototypes in the query set.

For usage instructions and code, please see the Github repo of this work: https://github.com/Brikwerk/MAE-FS

Model Date

November 2022

Model Type

MAE-FS uses a self-attention Transformer encoder and decoder for its architecture. CONV4, ResNet-18, and DINO-S are used a backbone models to decompose images into embedded representations. All weights provided are named according to the backbone model included in the respective weights file.

The following weights are provided:

• CONV4.pth (uses a CONV4 backbone with an embedding dimension of 512)
• CONV4_BASE.pth (uses a CONV4 backbone with an embedding dimension of 64)
• RESNET18.pth (uses a ResNet-18 backbone with an embedding dimension of 512)
• DINO.pth (uses a DINO-Small backbone with an embedding dimension of 384)