- Fully Connected Layer : This layer has full connections to the previous layer output and is computed by multiplying the input tensor with a learnt tensor of weights.
- ReLU\citealt{Krizhevsky2012} : A non-linearity introducing (or "Activation") operation. Specifically, Rectified Linear Unit (ReLU) performs the operation \(f\left(x\right)\ =\ \max\left(0,\ x\right)\) on the input tensor.
- Dropout\citealt{srivastava2014dropout} : Used to reduce overfitting in the model by sampling units from the training network to update while keeping the rest as it is - effectively "dropping out" of training iteration. At each iteration, a different sample is updated and hence a different sample is dropped, avoiding a potetntial over dependency on a single unit.
We extract features from the processed images using different pre-trained models and train corresponding shallow networks. We sample the dataset without replacement into buckets of 80%(training) & 20%(test). For the training process, we use Adam\citealt{kingma2014adam} optimization algorithm with the initial learning rate(\(\alpha\)) of 10-4. The training dataset is increased multiple folds by image augmentation as described in Section \ref{717120}. The trained model is evaluated on the validation data generated on the fly and is finally tested (reported results) on the independent test dataset.