Training Process

You can train deep learning neural networks for classification tasks by using methods such as training from scratch, or by transfer learning, or by feature extraction.

Training from Scratch

Training a deep learning neural network from scratch requires a large amount of labeled data. To create the network architecture by using Neural Network Toolbox™, you can use the built-in layers, define your own layers, or import layers from Caffe models. The neural network is then trained by using the large amounts of labeled data. Use trained network for predicting or classifying the unlabeled data. These networks can take few days or couple of weeks to train. Therefore, it is not a commonly used method for training networks.

For more information, see Get Started with Transfer Learning.

Transfer Learning

Transfer learning is used for cases where there is lack of labeled data. The existing network architectures, trained for scenarios with large amounts of labeled data, are used for this approach. The parameters of pretrained networks are modified to fit the unlabeled data. Therefore, transfer learning is used for transferring knowledge across various tasks. You can train or modify these networks faster so it is the most widely used training approach for deep learning applications.

For more information, see Get Started with Transfer Learning

Feature Extraction

Layers in deep learning networks are trained for extracting features from the input data. This approach uses the network as a feature extractor. The features extracted after the training process can be put into various machine learning models such as Support Vector Machines (SVM).