Python demo — a script that recomputes what this entry states and prints one line per check
A transformer is an artificial neural network (ANN) built by composing layers,
each of which transforms the features of a data point
made up of smaller units, so-called tokens, such as the
words of a text or the patches of an image. The feature vectors of the tokens are
stacked into a matrix; each layer maps this
matrix to one of the same shape, so that layers can
be composed freely into a deep ANN. The layers
alternate between token mixing, which combines information across
tokens, and a position-wise multilayer perceptron (MLP),
which transforms each token separately.
Definition
Many machine learning (ML) applications involve
data points that consist of smaller units, so-called
tokens: a text consists of words, and an image consists of
patches of pixels (see Fig. 1).
A transformer is an artificial neural network (ANN) that is built by composing
layers, each of which transforms the features of
such a token-based data point
(Vaswani et al., 2017).
B-riverThe feature vectors (or embeddings)
$\featurevec^{(1)}, \ldots, \featurevec^{(\contextlen)} \in \reals^{\featuredim}$
of the $\contextlen$ tokens are stacked into a matrix
$\mX = \big(\featurevec^{(1)}, \ldots, \featurevec^{(\contextlen)}\big)^{\top}
\in \reals^{\contextlen \times \featuredim}$. Each layer of
the original transformer maps such a matrix to a transformed
matrix of the same shape, so that layers can be
composed freely into a deep ANN
(Vaswani et al., 2017, Sect. 3.1). Some transformer variants
reduce the number of tokens in deeper layers
(Tay et al., 2022). For example, a transformer that
translates a sentence transforms the stacked word embeddings,
layer by layer, into feature vectors from which the
translated words are predicted (Vaswani et al., 2017).
Contemporary large language models (LLMs), such as GPT-3, are transformers trained
to predict the next token of a text (Brown et al., 2020);
transformers also process images by taking patches as
tokens (Dosovitskiy et al., 2021).
Figure 1: Two data points that consist of tokens.
Left: the opening words of the Universal Declaration of
Human Rights (Assembly, 1948); each word is one
token. Right: a weather-radar-based image of the
1-hour precipitation around Krems an der Donau, Austria
(August 20, 2026, 17:00 UTC; INCA analysis, GeoSphere
Austria), with $96 \times 96$ pixels of 1 km; darker
pixels indicate more precipitation, and thin lines cut
the image into $16 \times 16$-pixel patches, each of
which is one token. The course of the Danube and
the location of Krems are marked (river course:
OpenStreetMap contributors). Data generated by
pythondemos/transformer.py
The layers of a transformer alternate between two types of
transformations of $\mX$ (see Fig. 2).
A token-mixing layer maps $\mX$ to a new matrix
$\widetilde{\mX}$, each row of which is a weighted sum of the
rows of $\mX$, with weights that can depend on $\mX$ itself; the
tokens thereby interact. In the original transformer,
token mixing is implemented by multi-head attention
(Vaswani et al., 2017, Sect. 3.1); attention sets
transformers apart from recurrent neural networks (RNNs), which process the
tokens of a data point sequentially instead of in
parallel (Vaswani et al., 2017, Sect. 1). A
position-wise multilayer perceptron (MLP) then transforms each row
of $\widetilde{\mX}$ separately
(Vaswani et al., 2017, Sect. 3.3). Each of the two
transformations is wrapped in a residual connection, and a
subsequent layer normalization standardizes each row of the
resulting matrix (Vaswani et al., 2017, Sect. 3.1).
Figure 2: One transformer block. Each of the two
transformations, multi-head attention for token
mixing and a position-wise multilayer perceptron,
is bypassed by a residual connection (dashed) whose
output is added back and passed through
layer normalization. The largest Llama 3 LLM
stacks more than 120 of such blocks
(Grattafiori and others, 2024, Sect. 3.2)
The matrices produced by successive layers represent
the tokens at increasingly abstract levels. This
progression is measured by probing: a simple classifier is
trained to predict a chosen property from the output of a single
layer, and the depth at which the property becomes
predictable locates it in the network. Early layers of a
language transformer encode word-level properties, such as
part-of-speech tags, while relations between tokens far
apart in the text become predictable only in deeper layers
(Tenney et al., 2019).
Deep convolutional neural networks (CNNs) show the same
progression, from edges and textures in early layers to
object parts in deeper ones (Zeiler and Fergus, 2014), and vision
transformers develop it too: their early layers split
attention between nearby and distant patches, while the
deeper layers attend across the whole image
(Raghu et al., 2021). The most informative representations often
arise in the middle layers, whose outputs can outperform
those of the final layer as features for other tasks
(Skean et al., 2025).
Attention compares every pair of tokens, so its
cost grows as $\contextlen^{2}$ with the number $\contextlen$ of
tokens (Tay et al., 2022). This cost motivates
cheaper token mixing: any transformation that combines the rows
of $\mX$ can take the place of attention — a
multilayer perceptron applied across tokens
(Tolstikhin et al., 2021), a fixed Fourier transform
(Lee-Thorp et al., 2022), or a state-space layer
whose cost is linear in $\contextlen$ (Gu and Dao, 2024).
See also:attention, artificial neural network, layer, token,
embedding, word embedding, recurrent neural network, natural language processing,
large language model, multilayer perceptron,
residual connection, layer normalization,
convolutional neural network.
Dosovitskiy et al. (2021). An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale. Int. Conf. Learn. Represent. (ICLR). openreview.net/forum?id=YicbFdNTTy
Tenney et al. (2019). BERT Rediscovers the Classical NLP Pipeline. Proc. 57th Ann. Meeting Assoc. Comput. Linguistics. doi.org/10.18653/v1/P19-1452
Zeiler and Fergus (2014). Visualizing and Understanding Convolutional Networks. Proc. European Conf. Comput. Vis. (ECCV).
Raghu et al. (2021). Do Vision Transformers See Like Convolutional Neural Networks?. Adv. Neural Inf. Process. Syst..
Skean et al. (2025). Layer by Layer: Uncovering Hidden Representations in Language Models. Proc. Int. Conf. Mach. Learn. (ICML).
Tolstikhin et al. (2021). MLP-Mixer: An all-MLP Architecture for Vision. Adv. Neural Inf. Process. Syst.. doi.org/10.48550/arXiv.2105.01601
Lee-Thorp et al. (2022). FNet: Mixing Tokens with Fourier Transforms. Proc. Conf. North Amer. Chapter Assoc. Comput. Linguistics. doi.org/10.18653/v1/2022.naacl-main.319
Gu and Dao (2024). Mamba: Linear-Time Sequence Modeling with Selective State Spaces. Conf. Language Modeling. openreview.net/forum?id=tEYskw1VY2