The Wolfram Language makes it very easy to operate on neural networks using their symbolic representation. A new architecture can be built starting from another one by taking or adding layers, selecting and combining subgraphs or replacing specific parts or patterns. Networks can then be trained on a variety of data to be optimized for a specific task.

Network Training

NetTrain — train parameters in a net from examples

NetTrainResultsObject — represent what happened in a training

Network Operations

NetInitialize — randomly initialize parameters for a net

NetInsertSharedArrays — convert all arrays in a net into shared net arrays

Network Surgery

NetExtract — extract a specific layer or subgraph

NetUnfold — expose recurrent states of a net

NetReplacePart — replace layers or layer properties

NetFlatten — flatten nested net structures like subgraphs

NetJoin — combine a series of nets

NetRename — rename layers or subpart of a net

NetAppend, NetPrepend — add one or more layers before and after a net

NetTake  ▪  NetDrop  ▪  NetInsert  ▪  NetDelete  ▪  NetReplace

Network Composition

NetChain — chain composition of net layers

NetGraph — graph of net layers

Higher-Order Network Construction

NetMapOperator — map over a sequence

NetMapThreadOperator — map over multiple sequences

NetFoldOperator — recurrent network that folds in elements of a sequence

NetBidirectionalOperator — bidirectional recurrent network

NetNestOperator — apply the same operation multiple times

Special Training Operators

NetPairEmbeddingOperator — train a Siamese neural network

NetGANOperator — train generative adversarial networks (GAN)