Handwriting recognition using Convolutional neural network ad experiment against proving that CNN is better than deep neural networks in better classifying images with its improvised convolution and pooling functions using Keras on MNIST data set.
In my previous attempt to classify digits using Deep neural networks I obtained an accuracy of ~ 93% not so good on MNIST data set, But using Convolution neural network (CNN) I obtained a whooping accuracy of ~ 99% which proves that CNN is the best alternative to DNN for classifying images.
I have attached the images showing the plots of training error , and accuracy of both Deep neural network classification and Convolutional neural network classification
We can see here from the above image that the training error is comparitively less in CNN , though our model in DNN is trying to over fit the data as our validation set is performing better than our training set there. Epochs = 10
| Model | training set loss | validation set loss |
|---|---|---|
| DNN | ~0.22 | ~0.20 |
| CNN | ~0.05 | ~0.03 |
| Model | training accuracy | validation accuracy |
|---|---|---|
| DNN | ~93% | ~94% |
| CNN | ~98.7% | ~99% |