To classify images containing handwritten digits using multiple custom-built CNN architectures, which may or may not are inspired from standard Convnet architectures such as LeNet, AlexNet, VGGNet, ResNet etc. A comparison between performance of different architectures are done.
The purpose of this study is to try 3 drastically different Convnet Architectures on MNIST image database. The implementation is done in Keras.
-
Take the famous MNIST dataset as input. http://yann.lecun.com/exdb/mnist/
-
Feed it into 3-layered Convnet Architecture design inspired by LeNet, 1998 paper by LeCunn.
-
Find the accuracy and draw the Loss vs Epoch Plot.
-
Introduce Batch Normalization and Dropouts.
-
Evaluate the model again by estimating accuracy and drawing loss diagram.
-
Feed same input to 5 layered Convnet Architecture design inspired by VGGNet, 2014 paper by Andrew Zisserman.
-
Introduce Pooling, Dropouts & evaluate the model again.
-
Feed same input to 7 layered Convnet Architecture self-designed with different-sized filters & dense layers.
-
Introduce Batch Normalization and Dropouts & evaluate the model again.
-
Analyze the output from the above 3 architectures and draw conclusions.
The 3-layered architecture is different but inspired from the LeNet, 1998 paper by Le Cunn.
The 5-layered architecture is different but inspired from the VGGNet, 2014 paper by Andrew Zisserman.
The 7-layered Convolution Architecture is custom built with different kernel sizes and dropout/ max pool considerations.
-
The performance of standard-model inspired networks are found higher than complex custom built architectures.
-
The convergence of model M2 happened much before Model 1. Number of epochs required is less.
-
The 99.5% accuracy of VGGNet-inspired M2 model is better than LeNet-inspired M1.
-
The distribution of weights are found to be normally distributed.
-
The huge increase in number of filters and different sized kernels did not help much.
-
VGGNet-inspired 5-layered model, M2 is found to be model of choice. It even outperformed a 7-layered Convnet with huge number of parameters. The convergence speed w.r.t. epochs is also comparable between M2 and M3.