Shuffling labels in CIFAR10 dataset

[ckeshava]

Description

Reproduce the results of the experiment on shuffling the labels of the CIFAR10 dataset. More precisely, we need to shuffle the labels of the images in this dataset and verify if the training accuracy can indeed be reduced to zero.

Details

• Technical Specifications: tensorflow/pytorch
• Type of issue: Multiple
• Time Limit: 5 days

Issue requirements / progress

• Tensorflow: A graph of (decreasing) training-error versus number-of-iterations.
• Tensorflow: Values of parameters of the model.
• Pytorch: A graph of (decreasing) training-error versus number-of-iterations.
• Pytorch: Values of parameters of the model.

Resources

This is with respect to the research paper on "Understanding Deep Learning requires rethinking generalization"

Here is a link to the paper: https://arxiv.org/abs/1611.03530. Feel free to open new issues if you are interested to reproduce other experiments from this paper.

The Experimental Setup at the end of the paper is relevant for this task. Please use the values of the hyper-parameters as mentioned in the paper.

Directory Structure

Name the code file as <your_name>.py and save it inside /machine_learning/shuffling_cifar10_labels/pytorch/ck.py [for a pytorch implementation]

Note

Please claim the issue first by commenting here before starting to work on it.