Jack on RNN seq2seq models learning alignments at 1:30 pm (Feb 24)
Sequence-to-sequence (seq2seq) models are a class of neural network models that learn the mapping between given input and target sequences. These models take an encoder-decoder structure where the encoder reads the inputs and passes the encoded information to the decoder, and the decoder decodes the encoded information and writes outputs.
Jack's focus is on Recurrent Neural Networks (RNNs) seq2seq models, whose encoder and decoder are both RNNs, in learning four transduction tasks that require alignments learning. These tasks are: identity, reversal, total reduplication, and input specified reduplication. Jack will talk about the learning capabilities of RNN seq2seq models, and the complexity of learning them that is different from the traditional FST*-theoretic characterizations.
*FST: Finite State Transducer