Seq2Seq

Of all Neural Network approaches for this problem, a common denominator is the formulation of an encoder and a decoder: resulting in the alias "encoder-decoder networks"

1. Encoder(Seq2Seq)

  • maps sequential input to an embedding
    • serves as a machine-usable representation of the input sequence
  • may be a CNN, RNN, or some other architecture

2. Decoder(Seq2Seq)

  • maps the embedding to the desired output sequence

3. Misc

  • incorporating Attention modules into the encoder-decoder networks results in better performance
    • implemented by an additional set of parameters that combine some information from the encoder and the current state vector of the decoder to generate the label
    • results in better context consideration than bidirectional and gated RNN variants.
Tags::ml:ai: