Deep Learning Model Series start from 1998 to 2021
- The architecture consists of a total of 7 layers consisting- 2 sets of Convolution layers, and 2 sets of Average/Max pooling layers which are followed by a flattening convolution layer. After that, we have 2 dense fully connected layers and finally a softmax classifier.
- The input used by LeCun was of the size (32 x 32) but as we will be using the MNIST dataset, so the image size in this dataset is (28 x 28). Thus, the input size we’ll be having is (28 x 28).
- Accuraracy is 0.9810.
- Loss is 0.0602.
- Prediction correct.
- Works well for MNIST Dataset.
- Simple and less complex structured conv layer model
AlexNet proposed work has eight layers including five convolutional layers followed by three fully connected layers. Some of the convolutional layers of the model are followed by max-pooling layers. As an activation function, the ReLU function is used by the network which shows improved performance over sigmoid and tanh functions. Model Additional Features :
- Activation : Relu
- Dropout
- Local Response Normalization
- BatchNormalization is replace lrn
- Pooling
- Augmention
- GPU's
- Training accuracy: 0.9925
- Training loss: 0.0237
- Validation Accuracy score : 0.6277
- Validation Loss : 2.0386
- Predictions : Not as we expected as we might need more amount of data or more layer training.
- Work's good for memorization but not as good as generalization
- VGG16 is a convolution neural net (CNN ) architecture which was used to win ILSVR(Imagenet) competition in 2014.
- Most unique thing about VGG16 is that instead of having a large number of hyper-parameter they focused on having convolution layers of 3x3 filter with a stride 1 and always used same padding and maxpool layer of 2x2 filter of stride 2.
- VGG was discurved for reature learning and mostly used techique as of today. Model Additional Features :
- Simple and same kernals
- Parameters are fix such as stride = 1, padding = 'same' for all convolution layers.
- Max pooling, conv layers and dense layers are used in entire network
- For feature leraning mostly vgg network is used.
- Vgg 16 : 16 layers [13 conv ; 5 MaxPool : 3 Dense : 1 Softmax]
- Accuraracy is 0.8868. #Only Epoches
- Loss is 0.2587.
- Prediction correct.
- Works well for Cat-Dog Dataset.
- Simple and less complex structured conv layer model.
- Transfer Learning.