Noise-as-targets representation learning for cifar10. Implementation based on the arxiv-paper "Unsupervised Learning by Predicting Noise" by Bojanowski and Joulin: https://arxiv.org/abs/1704.05310
Training:
Get neighbors:
Current status: Freezed. Best cifar10 test classification accuracy after 50 epochs of unsupervised training: 43,8%, not clear how to chose parameters, discussions, feedback or suggestions are welcome!
Example results of nearest neighbor search on the learned representation (for Cifar 10 test examples):
First column: query images, second to sixth columns: nearest neighbors.