MGU

Project for Deep Learning course. Its aim was to check chosen methods tackling data imbalance.

Used methods

• Show more often images that were of imbalanced class.
• Modify weights for function loss to give higher weights to imbalanced classes.
• Generate additional images by using SMOTE.

Used Dataset:

We used CIFAR-10 dataset to first create three different imbalanced dataset scenarios and later train and test neural networks.

Created imbalanced strategies

They can be found in dataset_downloader.py

Used CNNs:

• ResNet101 from here
• Vgg16 from lecturer notebook.

Results

You can find them here

Links

To trained neural networks Link