CapsNet

CapsNet for MNIST Kannada dataset

Reference Paper - KANNADA-MNIST: A NEW HANDWRITTEN DIGITS DATASET FOR THE KANNADA LANGUAGE

There are two datasets mentioned in the paper - MNIST-10k-Test dataset and the Kannada-MNIST-Test dataset.

For the Kannada MNIST dataset, with 60, 000 − 10, 000 train-test split, CNNs achieved 97.13% top-1 accuracy. While the CapsNet achieves 98% accuracy.

The pre-trained CNN achieved 76.2% top-1 accuracy on the dig-10k dataset. While the CapsNet achieves 83.04% for the same.

CapsNet for Traffic Signal classification

Final test acc: 94.81

CIFAR dataset

Batch Size 10

N_epochs =5

acc = 66.53%

To Do

• Stack more convolutional layers before capsule layers.

• Increase the size of the capsule layers (more capsules, larger capsules etc.). Note that it may take a lot of time.

• Play with number of routing iterations in forward pass.

• Play with kernel size of convolutions in the first layer.

• Play with kernel size of capsules in the second layer.

• Try different variants of original implementation's loss function (change m+, m-, lambda).

• Try different loss functions (Hinge or pure MSE, or cross-entropy!).

• Try different weights for reconstruction loss.

• Data Preprocessing and Shuffle data.