From Wikipedia, the free encyclopedia
Jump to navigation Jump to search

Seq2seq is a family of machine learning approaches used for language processing.[1] Applications include language translation, image captioning, conversational models and text summarization.[2]


The algorithm was developed by Google for use in machine translation.[2]

In 2019, Facebook announced its use in symbolic integration and resolution of differential equations. The company claimed that it could solve complex equations more rapidly and with greater accuracy than commercial solutions such as Mathematica, MATLAB and Maple. First, the equation is parsed into a tree structure to avoid notational idiosyncrasies. An LSTM neural network then applies its standard pattern recognition facilities to process the tree.[3]

In 2020, Google released Meena, a 2.6 billion parameter seq2seq-based chatbot trained on a 341 GB data set. Google claimed that the chatbot has 1.7 times greater model capacity than OpenAI's GPT-2,[4] whose May 2020 successor, the 175 billion parameter GPT-3, trained on a "45TB dataset of plaintext words (45,000 GB) that was ... filtered down to 570 GB."[5]


Seq2seq turns one sequence into another sequence (sequence transformation). It does so by use of a recurrent neural network (RNN) or more often LSTM or GRU to avoid the problem of vanishing gradient. The context for each item is the output from the previous step. The primary components are one encoder and one decoder network. The encoder turns each item into a corresponding hidden vector containing the item and its context. The decoder reverses the process, turning the vector into an output item, using the previous output as the input context.[2]

Optimizations include:[2]

  • Attention: The input to the decoder is a single vector which stores the entire context. Attention allows the decoder to look at the input sequence selectively.
  • Beam Search: Instead of picking the single output (word) as the output, multiple highly probable choices are retained, structured as a tree (using a Softmax on the set of attention scores[6]). Average the encoder states weighted by the attention distribution.[6]
  • Bucketing: Variable-length sequences are possible because of padding with 0s, which may be done to both input and output. However, if the sequence length is 100 and the input is just 3 items long, expensive space is wasted. Buckets can be of varying sizes and specify both input and output lengths.

Training typically uses a cross-entropy loss function, whereby one output is penalized to the extent that the probability of the succeeding output is less than 1.[6]

Related software[edit]

Software adopting similar approaches includes OpenNMT (Torch), Neural Monkey (TensorFlow) and NEMATUS (Theano).[7]

See also[edit]


  1. ^ Sutskever, Ilya; Vinyals, Oriol; Le, Quoc Viet (2014). "Sequence to sequence learning with neural networks". arXiv:1409.3215 [cs.CL].
  2. ^ a b c d Wadhwa, Mani (2018-12-05). "seq2seq model in Machine Learning". GeeksforGeeks. Retrieved 2019-12-17.
  3. ^ "Facebook has a neural network that can do advanced math". MIT Technology Review. December 17, 2019. Retrieved 2019-12-17.
  4. ^ Mehta, Ivan (2020-01-29). "Google claims its new chatbot Meena is the best in the world". The Next Web. Retrieved 2020-02-03.
  5. ^ Gage, Justin. "What's GPT-3?". Retrieved August 1, 2020.
  6. ^ a b c Hewitt, John; Kriz, Reno (2018). "Sequence 2 sequence Models" (PDF). Stanford University.
  7. ^ "Overview - seq2seq". Retrieved 2019-12-17.

External links[edit]