-
Notifications
You must be signed in to change notification settings - Fork 13
2020 Interesting papers and ideas
Vinay Sisodia edited this page Jun 19, 2020
·
5 revisions
- My implementation of the core idea
-
Core idea: Learn (or use pre-trained model) unsupervised representations. Then use k-means to cluster images. For each image, take nearest
n
(=5ish) neighbors and train a model to classify. You want to maximize dot product of softmax out + entropy of label predictions.
- Thoughts: interesting idea but a long road to it being applied for non-research use.