With the help of siamese deep convolutional networks and the technique of few shot learning, I have created an AI model which predicts the similarity of two characters.
To save virtual address space, I have used only 15 percent of the data and yet achieved 98.7 percent validation accuracy on omniglot dataset with the help of undersampling technique.
The model was trained on Google Colab with the help of its Tesla K80 GPU.