Using Convnet as a feature extractor

---

• Using nolearn and lasagne for working with ConvNet

  • lasagne is based on Theano so GPU speedups will make a difference
  • nolearn library is a collection of utilities around the NN packages

• Working on MNIST dataset

• ConvNet architecture

  • 2 convolutional layers with pooling
  • 1 fully connected layer and the output layer
  • Dropouts between some layers, dropout set at 50%

• Prediction

  • visualize the confusion matrix

• Filter visualization

  • visualize layers of ConvNet

• Feature extraction

  • plotting the output layer activations
  • dense layer activations instead of forwarding to a classifier can be by themselves used as features on a linear classifier