Transformer (machine learning model)

A transformer is a deep learning model that adopts the mechanism of self-attention, differentially weighting the significance of each part of the input data. It is used primarily in the fields of natural language processing (NLP)[1] and computer vision (CV).[2]

Like recurrent neural networks (RNNs), transformers are designed to process sequential input data, such as natural language, with applications towards tasks such as translation and text summarization. However, unlike RNNs, transformers process the entire input all at once. The attention mechanism provides context for any position in the input sequence. For example, if the input data is a natural language sentence, the transformer does not have to process one word at a time. This allows for more parallelization than RNNs and therefore reduces training times.[1]

Transformers were introduced in 2017 by a team at Google Brain[1] and are increasingly the model of choice for NLP problems,[3] replacing RNN models such as long short-term memory (LSTM). The additional training parallelization allows training on larger datasets. This led to the development of pretrained systems such as BERT (Bidirectional Encoder Representations from Transformers) and GPT (Generative Pre-trained Transformer), which were trained with large language datasets, such as the Wikipedia Corpus and Common Crawl, and can be fine-tuned for specific tasks.[4][5]


Before transformers, most state-of-the-art NLP systems relied on gated RNNs, such as LSTMs and gated recurrent units (GRUs), with added attention mechanisms. Transformers also make use of attention mechanisms but, unlike RNNs, do not have a recurrent structure. This means that provided with enough training data, attention mechanisms alone can match the performance of RNNs with attention.[1]

Sequential processingEdit

Gated RNNs process tokens sequentially, maintaining a state vector that contains a representation of the data seen prior to the current token. To process the  th token, the model combines the state representing the sentence up to token   with the information of the new token to create a new state, representing the sentence up to token  . Theoretically, the information from one token can propagate arbitrarily far down the sequence, if at every point the state continues to encode contextual information about the token. In practice this mechanism is flawed: the vanishing gradient problem leaves the model's state at the end of a long sentence without precise, extractable information about preceding tokens. The dependency of token computations on the results of previous token computations also makes it hard to parallelize computation on modern deep-learning hardware. This can make the training of RNNs inefficient.


These problems were addressed by attention mechanisms. Attention mechanisms let a model draw from the state at any preceding point along the sequence. The attention layer can access all previous states and weigh them according to a learned measure of relevance, providing relevant information about far-away tokens.

A clear example of the value of attention is in language translation, where context is essential to assign the meaning of a word in a sentence. In an English-to-French translation system, the first word of the French output most probably depends heavily on the first few words of the English input. However, in a classic LSTM model, in order to produce the first word of the French output, the model is given only the state vector after processing the last English word. Theoretically, this vector can encode information about the whole English sentence, giving the model all the necessary knowledge. In practice, this information is often poorly preserved by the LSTM. An attention mechanism can be added to address this problem: the decoder is given access to the state vectors of every English input word, not just the last, and can learn attention weights that dictate how much to attend to each English input state vector.

When added to RNNs, attention mechanisms increase performance. The development of the Transformer architecture revealed that attention mechanisms were powerful in themselves and that sequential recurrent processing of data was not necessary to achieve the quality gains of RNNs with attention. Transformers use an attention mechanism without an RNN, processing all tokens simultaneously and calculating attention weights between them in successive layers. Since the attention mechanism only uses information about other tokens from lower layers, it can be computed for all tokens in parallel, which leads to improved training speed.


Transformer model architecture


The input text is parsed into tokens by a byte pair encoding tokenizer, and each token is converted via a word embedding into a vector. Then, positional information of the token is added to the word embedding.

Encoder–decoder architectureEdit

Like earlier seq2seq models, the original Transformer model used an encoder–decoder architecture. The encoder consists of encoding layers that process the input iteratively one layer after another, while the decoder consists of decoding layers that do the same thing to the encoder's output.

The function of each encoder layer is to generate encodings that contain information about which parts of the inputs are relevant to each other. It passes its encodings to the next encoder layer as inputs. Each decoder layer does the opposite, taking all the encodings and using their incorporated contextual information to generate an output sequence.[6] To achieve this, each encoder and decoder layer makes use of an attention mechanism.

For each part of the input, attention weighs the relevance of every other part and draws from them to produce the output.[7] Each decoder layer has an additional attention mechanism that draws information from the outputs of previous decoders, before the decoder layer draws information from the encodings.

Both the encoder and decoder layers have a feed-forward neural network for additional processing of the outputs and contain residual connections and layer normalization steps.[7]

Scaled dot-product attentionEdit

The transformer building blocks are scaled dot-product attention units. When a sentence is passed into a transformer model, attention weights are calculated between every token simultaneously. The attention unit produces embeddings for every token in context that contain information about the token itself along with a weighted combination of other relevant tokens each weighted by its attention weight.

For each attention unit, the transformer model learns three weight matrices; the query weights  , the key weights  , and the value weights  . For each token  , the input word embedding   is multiplied with each of the three weight matrices to produce a query vector  , a key vector  , and a value vector  . Attention weights are calculated using the query and key vectors: the attention weight   from token   to token   is the dot product between   and  . The attention weights are divided by the square root of the dimension of the key vectors,  , which stabilizes gradients during training, and passed through a softmax which normalizes the weights. The fact that   and   are different matrices allows attention to be non-symmetric: if token   attends to token   (i.e.   is large), this does not necessarily mean that token   will attend to token   (i.e.   could be small). The output of the attention unit for token   is the weighted sum of the value vectors of all tokens, weighted by  , the attention from token   to each token.

The attention calculation for all tokens can be expressed as one large matrix calculation using the softmax function, which is useful for training due to computational matrix operation optimizations that quickly compute matrix operations. The matrices  ,   and   are defined as the matrices where the  th rows are vectors  ,  , and   respectively. Then we can represent the attention as


where softmax is taken over the horizontal axis.

Multi-head attentionEdit

One set of   matrices is called an attention head, and each layer in a transformer model has multiple attention heads. While each attention head attends to the tokens that are relevant to each token, with multiple attention heads the model can do this for different definitions of "relevance". In addition, the influence field representing relevance can become progressively dilated in successive layers. Many transformer attention heads encode relevance relations that are meaningful to humans. For example, some attention heads can attend mostly to the next word, while others mainly attend from verbs to their direct objects.[8] The computations for each attention head can be performed in parallel, which allows for fast processing. The outputs for the attention layer are concatenated to pass into the feed-forward neural network layers.

Concretely, let the multiple attention heads be indexed by  , then we have

where the matrices   are "projection matrices" owned by individual attention head  , and   is a final projection matrix owned by the whole multi-headed attention head.

Masked attentionEdit

It may be necessary to cut out attention links between some word-pairs. For example, the decoder for token position   should not have access to token position  . This may be accomplished before the softmax stage by adding a mask matrix   that is negative infinity at entries where the attention link must be cut, and zero at other places.


Each encoder consists of two major components: a self-attention mechanism and a feed-forward neural network. The self-attention mechanism accepts input encodings from the previous encoder and weights their relevance to each other to generate output encodings. The feed-forward neural network further processes each output encoding individually. These output encodings are then passed to the next encoder as its input, as well as to the decoders.

The first encoder takes positional information and embeddings of the input sequence as its input, rather than encodings. The positional information is necessary for the transformer to make use of the order of the sequence, because no other part of the transformer makes use of this.[1]

The encoder is bidirectional. Attention can be placed on tokens before and after the current token. Tokens are used instead of words to account for polysemy.

A diagram of a sinusoidal positional encoding with parameters  

Positional encodingEdit

A positional encoding is a fixed-size vector representation that encapsulates the relative positions of tokens within a target sequence: it provides the transformer model with information about where the words are in the input sequence.

The positional encoding is defined as a function of type  , where   is a positive even integer. The full positional encoding - as defined in the original paper - is given by the equation:

where  .

Here,   is a free parameter that should be significantly larger than the biggest   that would be input into the positional encoding function. In the original paper,[1] the authors chose  .

The function is in a simpler form when written as a complex function of type  

where  .

The main reason the authors chose this as the positional encoding function is that it allows one to perform shifts as linear transformations:

where   is the distance one wishes to shift. This allows the transformer to take any encoded position, and find the encoding of the position n-steps-ahead or n-steps-behind, by a matrix multiplication.

By taking a linear sum, any convolution can also be implemented as linear transformations:

for any constants  . This allows the transformer to take any encoded position and find a linear sum of the encoded locations of its neighbors. This sum of encoded positions, when fed into the attention mechanism, would create attention weights on its neighbors, much like what happens in a convolutional neural network language model. In the author's words, "we hypothesized it would allow the model to easily learn to attend by relative position".

In typical implementations, all operations are done over the real numbers, not the complex numbers, but since complex multiplication can be implemented as real 2-by-2 matrix multiplication, this is a mere notational difference.

Other positional encoding schemes exist.[9]


Each decoder consists of three major components: a self-attention mechanism, an attention mechanism over the encodings, and a feed-forward neural network. The decoder functions in a similar fashion to the encoder, but an additional attention mechanism is inserted which instead draws relevant information from the encodings generated by the encoders. This mechanism can also be called the encoder-decoder attention.[1][7]

Like the first encoder, the first decoder takes positional information and embeddings of the output sequence as its input, rather than encodings. The transformer must not use the current or future output to predict an output, so the output sequence must be partially masked to prevent this reverse information flow.[1] This allows for autoregressive text generation. For all attention heads, attention can't be placed on following tokens. The last decoder is followed by a final linear transformation and softmax layer, to produce the output probabilities over the vocabulary.

GPT has a decoder-only architecture.


Training transformer-based architectures can be expensive, especially for long inputs.[10] Alternative architectures include the Reformer (which reduces the computational load from   to  [10]), or models like ETC/BigBird (which can reduce it to  )[11] where   is the length of the sequence. This is done using locality-sensitive hashing and reversible layers.[12][13]

Ordinary transformers require a memory size that is quadratic in the size of the context window. Attention Free Transformers[14] reduce this to a linear dependence while still retaining the advantages of a transformer by linking the key to the value.

A benchmark for comparing transformer architectures was introduced in late 2020 by the name of Long Range Arena.[15]


Methods for stabilizing trainingEdit

The plain Transformer architecture has difficulty converging. In the original paper[1] the authors recommended using learning rate warmup. That is, the learning rate should linearly scale up from 0 to maximal value for the first part of the training (usually recommended to be 2% of the total number of training steps), before decaying again.

[16] found that using layer normalization before (instead of after) multiheaded attention and feedforward layers stabilizes training, not requiring learning rate warmup.


Transformers typically undergo self-supervised learning involving unsupervised pretraining followed by supervised fine-tuning. Pretraining is typically done on a larger dataset than fine-tuning, due to the limited availability of labeled training data. Tasks for pretraining and fine-tuning commonly include:


The transformer has had great success in natural language processing (NLP), for example the tasks of machine translation and time series prediction. Many pretrained models such as GPT-2, GPT-3, GPT-4, BERT, XLNet, RoBERTa and ChatGPT demonstrate the ability of transformers to perform a wide variety of such NLP-related tasks, and have the potential to find real-world applications. These may include:


The transformer model has been implemented in standard deep learning frameworks such as TensorFlow and PyTorch.

Transformers is a library produced by Hugging Face that supplies transformer-based architectures and pretrained models.[3]

See alsoEdit


  1. ^ a b c d e f g h i Vaswani, Ashish; Shazeer, Noam; Parmar, Niki; Uszkoreit, Jakob; Jones, Llion; Gomez, Aidan N.; Kaiser, Lukasz; Polosukhin, Illia (2017-06-12). "Attention Is All You Need". arXiv:1706.03762 [cs.CL].
  2. ^ He, Cheng (31 December 2021). "Transformer in CV". Transformer in CV. Towards Data Science.
  3. ^ a b Wolf, Thomas; Debut, Lysandre; Sanh, Victor; Chaumond, Julien; Delangue, Clement; Moi, Anthony; Cistac, Pierric; Rault, Tim; Louf, Remi; Funtowicz, Morgan; Davison, Joe; Shleifer, Sam; von Platen, Patrick; Ma, Clara; Jernite, Yacine; Plu, Julien; Xu, Canwen; Le Scao, Teven; Gugger, Sylvain; Drame, Mariama; Lhoest, Quentin; Rush, Alexander (2020). "Transformers: State-of-the-Art Natural Language Processing". Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing: System Demonstrations. pp. 38–45. doi:10.18653/v1/2020.emnlp-demos.6. S2CID 208117506.
  4. ^ a b c "Open Sourcing BERT: State-of-the-Art Pre-training for Natural Language Processing". Google AI Blog. Retrieved 2019-08-25.
  5. ^ a b "Better Language Models and Their Implications". OpenAI. 2019-02-14. Retrieved 2019-08-25.
  6. ^ "Sequence Modeling with Neural Networks (Part 2): Attention Models". Indico. 2016-04-18. Retrieved 2019-10-15.
  7. ^ a b c Alammar, Jay. "The Illustrated Transformer". Retrieved 2019-10-15.
  8. ^ Clark, Kevin; Khandelwal, Urvashi; Levy, Omer; Manning, Christopher D. (August 2019). "What Does BERT Look at? An Analysis of BERT's Attention". Proceedings of the 2019 ACL Workshop BlackboxNLP: Analyzing and Interpreting Neural Networks for NLP. Florence, Italy: Association for Computational Linguistics: 276–286. doi:10.18653/v1/W19-4828.
  9. ^ Dufter, Philipp; Schmitt, Martin; Schütze, Hinrich (2022-06-06). "Position Information in Transformers: An Overview". Computational Linguistics. 48 (3): 733–763. doi:10.1162/coli_a_00445. ISSN 0891-2017. S2CID 231986066.
  10. ^ a b Kitaev, Nikita; Kaiser, Łukasz; Levskaya, Anselm (2020). "Reformer: The Efficient Transformer". arXiv:2001.04451 [cs.LG].
  11. ^ "Constructing Transformers For Longer Sequences with Sparse Attention Methods". Google AI Blog. Retrieved 2021-05-28.
  12. ^ "Tasks with Long Sequences – Chatbot". Coursera.
  13. ^ "Reformer: The Efficient Transformer". Google AI Blog. Retrieved 2020-10-22.
  14. ^ Zhai, Shuangfei; Talbott, Walter; Srivastava, Nitish; Huang, Chen; Goh, Hanlin; Zhang, Ruixiang; Susskind, Josh (2021-09-21). "An Attention Free Transformer". arXiv:2105.14103 [cs.LG].
  15. ^ Tay, Yi; Dehghani, Mostafa; Abnar, Samira; Shen, Yikang; Bahri, Dara; Pham, Philip; Rao, Jinfeng; Yang, Liu; Ruder, Sebastian; Metzler, Donald (2020-11-08). "Long Range Arena: A Benchmark for Efficient Transformers". arXiv:2011.04006 [cs.LG].
  16. ^ Xiong, Ruibin; Yang, Yunchang; He, Di; Zheng, Kai; Zheng, Shuxin; Xing, Chen; Zhang, Huishuai; Lan, Yanyan; Wang, Liwei; Liu, Tie-Yan (2020-06-29). "On Layer Normalization in the Transformer Architecture". arXiv:2002.04745 [cs, stat].
  17. ^ a b Wang, Alex; Singh, Amanpreet; Michael, Julian; Hill, Felix; Levy, Omer; Bowman, Samuel (2018). "GLUE: A Multi-Task Benchmark and Analysis Platform for Natural Language Understanding". Proceedings of the 2018 EMNLP Workshop BlackboxNLP: Analyzing and Interpreting Neural Networks for NLP. Stroudsburg, PA, USA: Association for Computational Linguistics: 353–355. arXiv:1804.07461. doi:10.18653/v1/w18-5446. S2CID 5034059.

Further readingEdit

– Discussion of the effect of a transformer layer as equivalent to a Hopfield update, bringing the input closer to one of the fixed points (representable patterns) of a continuous-valued Hopfield network