What other approaches can be used to initialize weights for training fully connected deep architectures, given that greedy layer-wise pretraining is not required?

One alternative approach is joint learning with random initialization. Another option is using transfer learning from a pre-trained model on a related task.

Tagged 'joint_learning', 'transfer_learning', ['neural_networks'

What other approaches can be used to initialize weights for training fully connected deep architectures, given that greedy layer-wise pretraining is not required?

Leave a Reply Cancel reply