Backpropagation

As a machine-learning algorithm, backpropagation performs a backward pass to adjust the model's parameters, aiming to minimize the mean squared error
(MSE).[1][2] In a single-layered network, backpropagation uses the following steps:

1. Traverse through the network from the input to the output by computing the hidden layers' output and the output layer. (the feedforward step)
2. In the output layer, calculate the derivative of the cost function with respect to the input and the hidden layers.
3. Repeatedly update the weights until they converge or the model has undergone enough iterations.

It is an efficient application of the Leibniz chain rule (1673)[3] to such networks.[4] It is also known as the reverse mode of automatic differentiation or reverse
accumulation, due to Seppo Linnainmaa (1970).[5][6][7][8][9][10][11] The term "back-propagating error correction" was introduced in 1962 by Frank
Rosenblatt,[12][4] but he did not know how to implement this, even though Henry J. Kelley had a continuous precursor of backpropagation[13] already in 1960 in
the context of control theory.[4]

Backpropagation computes the gradient of a loss function with respect to the weights of the network for a single input–output example, and does so efficiently,
computing the gradient one layer at a time, iterating backward from the last layer to avoid redundant calculations of intermediate terms in the chain rule; this can be
derived through dynamic programming.[13][14][15] Gradient descent, or variants such as stochastic gradient descent,[16] are commonly used.

Strictly the term backpropagation refers only to the algorithm for computing the gradient, not how the gradient is used; but the term is often used loosely to refer to
the entire learning algorithm – including how the gradient is used, such as by stochastic gradient descent.[17] In 1986 David E. Rumelhart et al. published an
experimental analysis of the technique.[18] This contributed to the popularization of backpropagation and helped to initiate an active period of research in
multilayer perceptrons.

Overview
Backpropagation computes the gradient in weight space of a feedforward neural network, with respect to a loss function. Denote:

: input (vector of features)

: target output

For classification, output will be a vector of class probabilities (e.g., , and target output is a specific class, encoded by the
one-hot/dummy variable (e.g., ).

: loss function or "cost function"[a]

For classification, this is usually cross entropy (XC, log loss), while for regression it is usually squared error loss (SEL).

: the number of layers

: the weights between layer and , where is the weight between the -th node in layer and the -th node in layer [b]
: activation functions at layer

For classification the last layer is usually the logistic function for binary classification, and softmax (softargmax) for multi-class
classification, while for the hidden layers this was traditionally a sigmoid function (logistic function or others) on each node (coordinate),
but today is more varied, with rectifier (ramp, ReLU) being common.

: activation of the -th node in layer .

In the derivation of backpropagation, other intermediate quantities are used by introducing them as needed below. Bias terms are not treated specially since they
correspond to a weight with a fixed input of 1. For backpropagation the specific loss function and activation functions do not matter as long as they and their
derivatives can be evaluated efficiently. Traditional activation functions include sigmoid, tanh, and ReLU. swish[19] mish,[20] and other activation functions have
since been proposed as well.

The overall network is a combination of function composition and matrix multiplication:

For a training set there will be a set of input–output pairs, . For each input–output pair in the training set, the loss of the model on that pair is the
cost of the difference between the predicted output and the target output :

Note the distinction: during model evaluation the weights are fixed while the inputs vary (and the target output may be unknown), and the network ends with the
output layer (it does not include the loss function). During model training the input–output pair is fixed while the weights vary, and the network ends with the loss
function.

Backpropagation computes the gradient for a fixed input–output pair , where the weights can vary. Each individual component of the gradient,
can be computed by the chain rule; but doing this separately for each weight is inefficient. Backpropagation efficiently computes the gradient by
avoiding duplicate calculations and not computing unnecessary intermediate values, by computing the gradient of each layer – specifically the gradient of the
weighted input of each layer, denoted by – from back to front.
Informally, the key point is that since the only way a weight in affects the loss is through its effect on the next layer, and it does so linearly, are the only data
you need to compute the gradients of the weights at layer , and then the previous layer can be computed and repeated recursively. This avoids inefficiency in
two ways. First, it avoids duplication because when computing the gradient at layer – it is unnecessary to recompute all derivatives on later layers
each time. Second, it avoids unnecessary intermediate calculations, because at each stage it directly computes the gradient of the weights with
respect to the ultimate output (the loss), rather than unnecessarily computing the derivatives of the values of hidden layers with respect to changes in weights
.

Backpropagation can be expressed for simple feedforward networks in terms of matrix multiplication, or more generally in terms of the adjoint graph.

Matrix multiplication
For the basic case of a feedforward network, where nodes in each layer are connected only to nodes in the immediate next layer (without skipping any layers), and
there is a loss function that computes a scalar loss for the final output, backpropagation can be understood simply by matrix multiplication.[c] Essentially,
backpropagation evaluates the expression for the derivative of the cost function as a product of derivatives between each layer from right to left – "backwards" –
with the gradient of the weights between each layer being a simple modification of the partial products (the "backwards propagated error").

Given an input–output pair , the loss is:

To compute this, one starts with the input and works forward; denote the weighted input of each hidden layer as and the output of hidden layer as the
activation . For backpropagation, the activation as well as the derivatives (evaluated at ) must be cached for use during the backwards pass.

The derivative of the loss in terms of the inputs is given by the chain rule; note that each term is a total derivative, evaluated at the value of the network (at each
node) on the input :

where is a diagonal matrix.

These terms are: the derivative of the loss function;[d] the derivatives of the activation functions;[e] and the matrices of weights:[f]

The gradient is the transpose of the derivative of the output in terms of the input, so the matrices are transposed and the order of multiplication is reversed, but
the entries are the same:

Backpropagation then consists essentially of evaluating this expression from right to left (equivalently, multiplying the previous expression for the derivative from
left to right), computing the gradient at each layer on the way; there is an added step, because the gradient of the weights isn't just a subexpression: there's an extra
multiplication.

Introducing the auxiliary quantity for the partial products (multiplying from right to left), interpreted as the "error at level " and defined as the gradient of the
input values at level :

Note that is a vector, of length equal to the number of nodes in level ; each component is interpreted as the "cost attributable to (the value of) that node".

The gradient of the weights in layer is then:

The factor of is because the weights between level and affect level proportionally to the inputs (activations): the inputs are fixed, the weights
vary.

The can easily be computed recursively, going from right to left, as:

The gradients of the weights can thus be computed using a few matrix multiplications for each level; this is backpropagation.

Compared with naively computing forwards (using the for illustration):

there are two key differences with backpropagation:

 1. Computing in terms of avoids the obvious duplicate multiplication of layers and beyond.
2. Multiplying starting from – propagating the error backwards – means that each step simply multiplies a vector ( ) by the matrices of
weights and derivatives of activations . By contrast, multiplying forwards, starting from the changes at an earlier layer, means
that each multiplication multiplies a matrix by a matrix. This is much more expensive, and corresponds to tracking every possible path of a
change in one layer forward to changes in the layer (for multiplying by , with additional multiplications for the derivatives of
the activations), which unnecessarily computes the intermediate quantities of how weight changes affect the values of hidden nodes.

Adjoint graph
For more general graphs, and other advanced variations, backpropagation can be understood in terms of automatic differentiation, where backpropagation is a
special case of reverse accumulation (or "reverse mode").[11]

Intuition

Motivation

The goal of any supervised learning algorithm is to find a function that best maps a set of inputs to their correct output. The motivation for backpropagation is to
train a multi-layered neural network such that it can learn the appropriate internal representations to allow it to learn any arbitrary mapping of input to output.[21]

Learning as an optimization problem

To understand the mathematical derivation of the backpropagation algorithm, it helps to first develop some intuition about the relationship between the actual
output of a neuron and the correct output for a particular training example. Consider a simple neural network with two input units, one output unit and no hidden
units, and in which each neuron uses a linear output (unlike most work on neural networks, in which mapping from inputs to outputs is non-linear)[g] that is the
weighted sum of its input.

Initially, before training, the weights will be set randomly. Then the neuron learns from training examples, which in
this case consist of a set of tuples where and are the inputs to the network and t is the correct output
(the output the network should produce given those inputs, when it has been trained). The initial network, given
and , will compute an output y that likely differs from t (given random weights). A loss function is used for
measuring the discrepancy between the target output t and the computed output y. For regression analysis problems
the squared error can be used as a loss function, for classification the categorical crossentropy can be used.

As an example consider a regression problem using the square error as a loss:

where E is the discrepancy or error.

Consider the network on a single training case: . Thus, the input and are 1 and 1 respectively and the A simple neural network with two input
units (each with a single input) and one
correct output, t is 0. Now if the relation is plotted between the network's output y on the horizontal axis and the error
output unit (with two inputs)
E on the vertical axis, the result is a parabola. The minimum of the parabola corresponds to the output y which
minimizes the error E. For a single training case, the minimum also touches the horizontal axis, which means the error
will be zero and the network can produce an output y that exactly matches the target output t. Therefore, the problem
of mapping inputs to outputs can be reduced to an optimization problem of finding a function that will produce the minimal error.

However, the output of a neuron depends on the weighted sum of all its inputs:

where and are the weights on the connection from the input units to the output unit. Therefore, the error also
depends on the incoming weights to the neuron, which is ultimately what needs to be changed in the network to
enable learning.

In this example, upon injecting the training data , the loss function becomes

Error surface of a linear neuron for a

Then, the loss function takes the form of a parabolic cylinder with its base directed along . Since all sets
single training case
of weights that satisfy minimize the loss function, in this case additional constraints are required to
converge to a unique solution. Additional constraints could either be generated by setting specific conditions to the
weights, or by injecting additional training data.

One commonly used algorithm to find the set of weights that minimizes the error is gradient descent. By backpropagation, the steepest descent direction is
calculated of the loss function versus the present synaptic weights. Then, the weights can be modified along the steepest descent direction, and the error is
minimized in an efficient way.

Derivation
The gradient descent method involves calculating the derivative of the loss function with respect to the weights of the network. This is normally done using
backpropagation. Assuming one output neuron,[h] the squared error function is
where

is the loss for the output and target value ,

is the target output for a training sample, and
is the actual output of the output neuron.

For each neuron , its output is defined as

where the activation function is non-linear and differentiable over the activation region (the ReLU is not differentiable at one point). A historically used
activation function is the logistic function:

which has a convenient derivative of:

The input to a neuron is the weighted sum of outputs of previous neurons. If the neuron is in the first layer after the input layer, the of the input layer
are simply the inputs to the network. The number of input units to the neuron is . The variable denotes the weight between neuron of the previous layer
and neuron of the current layer.

Finding the derivative of the error

Calculating the partial derivative of the error with respect to a weight is done using the
chain rule twice:

Diagram of an artificial neural network to illustrate the notation used

here

 
 

 
 
(Eq. 1)

In the last factor of the right-hand side of the above, only one term in the sum depends on , so that

 
 

   
(Eq. 2)

If the neuron is in the first layer after the input layer, is just .

The derivative of the output of neuron with respect to its input is simply the partial derivative of the activation function:

 
 

   
(Eq. 3)

which for the logistic activation function

This is the reason why backpropagation requires that the activation function be differentiable. (Nevertheless, the ReLU activation function, which is non-
differentiable at 0, has become quite popular, e.g. in AlexNet)

The first factor is straightforward to evaluate if the neuron is in the output layer, because then and

 
 

 
 
(Eq. 4)

If half of the square error is used as loss function we can rewrite it as

However, if is in an arbitrary inner layer of the network, finding the derivative with respect to is less obvious.
Considering as a function with the inputs being all neurons receiving input from neuron ,

and taking the total derivative with respect to , a recursive expression for the derivative is obtained:

     

Therefore, the derivative with respect to can be calculated if all the derivatives with respect to the outputs of the next layer – the ones closer to the output
neuron – are known. [Note, if any of the neurons in set were not connected to neuron , they would be independent of and the corresponding partial
derivative under the summation would vanish to 0.]

Substituting Eq. 2, Eq. 3 Eq.4 and Eq. 5 in Eq. 1 we obtain:

with

if is the logistic function, and the error is the square error:

To update the weight using gradient descent, one must choose a learning rate, . The change in weight needs to reflect the impact on of an increase or

decrease in . If , an increase in increases ; conversely, if , an increase in decreases . The new is added to the old

weight, and the product of the learning rate and the gradient, multiplied by guarantees that changes in a way that always decreases . In other words, in

the equation immediately below, always changes in such a way that is decreased:

Second-order gradient descent

Using a Hessian matrix of second-order derivatives of the error function, the Levenberg-Marquardt algorithm often converges faster than first-order gradient
descent, especially when the topology of the error function is complicated.[22][23] It may also find solutions in smaller node counts for which other methods might
not converge.[23] The Hessian can be approximated by the Fisher information matrix.[24]

Loss function
The loss function is a function that maps values of one or more variables onto a real number intuitively representing some "cost" associated with those values. For
backpropagation, the loss function calculates the difference between the network output and its expected output, after a training example has propagated through
the network.

Assumptions

The mathematical expression of the loss function must fulfill two conditions in order for it to be possibly used in backpropagation.[25] The first is that it can be
written as an average over error functions , for individual training examples, . The reason for this assumption is that the backpropagation
algorithm calculates the gradient of the error function for a single training example, which needs to be generalized to the overall error function. The second
assumption is that it can be written as a function of the outputs from the neural network.

Example loss function

Let be vectors in .

Select an error function measuring the difference between two outputs. The standard choice is the square of the Euclidean distance between the vectors
and :

The error function over training examples can then be written as an average of losses over individual examples:
Limitations
Gradient descent with backpropagation is not guaranteed to find the global minimum of the error
function, but only a local minimum; also, it has trouble crossing plateaus in the error function
landscape. This issue, caused by the non-convexity of error functions in neural networks, was long
thought to be a major drawback, but Yann LeCun et al. argue that in many practical problems, it is
not.[26]
Backpropagation learning does not require normalization of input vectors; however, normalization
could improve performance.[27]
Backpropagation requires the derivatives of activation functions to be known at network design time.

History
Gradient descent may find a local
Modern backpropagation is Seppo Linnainmaa's reverse mode of automatic differentiation (1970) for discrete minimum instead of the global minimum.
connected networks of nested differentiable functions.[5][6][9][10][7][8] It is an efficient application of the chain rule
(derived by Gottfried Wilhelm Leibniz in 1673[3][28]) to such networks.[4] The terminology "back-propagating error
correction" was introduced in 1962 by Frank Rosenblatt,[29][4] but he did not know how to implement this, although Henry J. Kelley had a continuous precursor
of backpropagation[13] already in 1960 in the context of control theory.[4] The first deep learning multilayer perceptron (MLP) trained by stochastic gradient
descent[16] was published in 1967 by Shun'ichi Amari.[30][4] In computer experiments, his five layer MLP with two modifiable layers learned internal
representations required to classify non-linearily separable pattern classes.[4] In 1982, Paul Werbos applied backpropagation to MLPs in the way that has become
standard.[31][32][4] In 1985, David E. Rumelhart et al. published an experimental analysis of the technique.[18] This contributed to the popularization of
backpropagation and helped to initiate an active period of research in multilayer perceptrons.[21][33][34]

Kelley (1960)[13] and Arthur E. Bryson (1961)[14] used principles of dynamic programming to derive the above-mentioned continuous precursor of the method. In
1962, Stuart Dreyfus published a simpler derivation based only on the chain rule.[35][36][37][9][10] In 1973, he adapted parameters of controllers in proportion to
error gradients.[38] Unlike Linnainmaa's 1970 method,[5][7] these precursors used "standard Jacobian matrix calculations from one stage to the previous one,
neither addressing direct links across several stages nor potential additional efficiency gains due to network sparsity."[4]

In 1985, the method was also described by Parker.[39][40] Yann LeCun proposed an alternative form of backpropagation for neural networks in his PhD thesis in
1987. In 1993, Eric Wan won an international pattern recognition contest through backpropagation.[9][41]

During the 2000s it fell out of favour, but returned in the 2010s, benefitting from cheap, powerful GPU-based computing systems. This has been especially so in
speech recognition, machine vision, natural language processing, and language structure learning research (in which it has been used to explain a variety of
phenomena related to first[42] and second language learning.[43]).

Error backpropagation has been suggested to explain human brain ERP components like the N400 and P600.[44]

In 2023, a backpropagation algorithm was implemented on a photonic processor by a team at Stanford University.[45]

See also
Artificial neural network
Neural circuit
Catastrophic interference
Ensemble learning
AdaBoost
Overfitting
Neural backpropagation
Backpropagation through time

Notes
a. Use for the loss function to allow to be used for the number of layers
b. This follows Nielsen (2015), and means (left) multiplication by the matrix corresponds to converting output values of layer to input
values of layer : columns correspond to input coordinates, rows correspond to output coordinates.
c. This section largely follows and summarizes Nielsen (2015).
d. The derivative of the loss function is a covector, since the loss function is a scalar-valued function of several variables.
e. The activation function is applied to each node separately, so the derivative is just the diagonal matrix of the derivative on each node. This is
often represented as the Hadamard product with the vector of derivatives, denoted by , which is mathematically identical but better
matches the internal representation of the derivatives as a vector, rather than a diagonal matrix.
f. Since matrix multiplication is linear, the derivative of multiplying by a matrix is just the matrix: .
g. One may notice that multi-layer neural networks use non-linear activation functions, so an example with linear neurons seems obscure.
However, even though the error surface of multi-layer networks are much more complicated, locally they can be approximated by a
paraboloid. Therefore, linear neurons are used for simplicity and easier understanding.
 h. There can be multiple output neurons, in which case the error is the squared norm of the difference vector.

References
1. Goodfellow, Bengio & Courville 2016, p. 200 (https://www.deeplear 16. Robbins, H.; Monro, S. (1951). "A Stochastic Approximation
ningbook.org/contents/mlp.html#pf25), "Furthermore, back- Method" (https://doi.org/10.1214%2Faoms%2F1177729586). The
propagation is often misunderstood as being specific to multi-layer Annals of Mathematical Statistics. 22 (3): 400.
neural networks, but in principle it can compute derivatives of any doi:10.1214/aoms/1177729586 (https://doi.org/10.1214%2Faoms%
function" 2F1177729586).
2. Graves, Alex; Wayne, Greg; Danihelka, Ivo (2014). "Neural Turing 17. Goodfellow, Bengio & Courville 2016, p. 200 (https://www.deeplear
Machines". arXiv:1410.5401 (https://arxiv.org/abs/1410.5401) ningbook.org/contents/mlp.html#pf25), "The term back-propagation
[cs.NE (https://arxiv.org/archive/cs.NE)]. is often misunderstood as meaning the whole learning algorithm for
3. Leibniz, Gottfried Wilhelm Freiherr von (1920). The Early multilayer neural networks. Backpropagation refers only to the
Mathematical Manuscripts of Leibniz: Translated from the Latin method for computing the gradient, while other algorithms, such as
Texts Published by Carl Immanuel Gerhardt with Critical and stochastic gradient descent, is used to perform learning using this
Historical Notes (Leibniz published the chain rule in a 1676 gradient."
memoir) (https://books.google.com/books?id=bOIGAAAAYAAJ&q=l 18. Rumelhart; Hinton; Williams (1986). "Learning representations by
eibniz+altered+manuscripts&pg=PA90). Open court publishing back-propagating errors" (http://www.cs.toronto.edu/~hinton/absps/n
Company. aturebp.pdf) (PDF). Nature. 323 (6088): 533–536.
4. Schmidhuber, Juergen (2022). "Annotated History of Modern AI and Bibcode:1986Natur.323..533R (https://ui.adsabs.harvard.edu/abs/1
Deep Learning". arXiv:2212.11279 (https://arxiv.org/abs/2212.1127 986Natur.323..533R). doi:10.1038/323533a0 (https://doi.org/10.103
9) [cs.NE (https://arxiv.org/archive/cs.NE)]. 8%2F323533a0). S2CID 205001834 (https://api.semanticscholar.or
g/CorpusID:205001834).
5. Linnainmaa, Seppo (1970). The representation of the cumulative
rounding error of an algorithm as a Taylor expansion of the local 19. Ramachandran, Prajit; Zoph, Barret; Le, Quoc V. (2017-10-27).
rounding errors (Masters) (in Finnish). University of Helsinki. pp. 6– "Searching for Activation Functions". arXiv:1710.05941 (https://arxi
7. v.org/abs/1710.05941) [cs.NE (https://arxiv.org/archive/cs.NE)].
6. Linnainmaa, Seppo (1976). "Taylor expansion of the accumulated 20. Misra, Diganta (2019-08-23). "Mish: A Self Regularized Non-
rounding error". BIT Numerical Mathematics. 16 (2): 146–160. Monotonic Activation Function". arXiv:1908.08681 (https://arxiv.org/
doi:10.1007/bf01931367 (https://doi.org/10.1007%2Fbf01931367). abs/1908.08681) [cs.LG (https://arxiv.org/archive/cs.LG)].
S2CID 122357351 (https://api.semanticscholar.org/CorpusID:12235 21. Rumelhart, David E.; Hinton, Geoffrey E.; Williams, Ronald J.
7351). (1986a). "Learning representations by back-propagating errors".
7. Griewank, Andreas (2012). "Who Invented the Reverse Mode of Nature. 323 (6088): 533–536. Bibcode:1986Natur.323..533R (http
Differentiation?". Optimization Stories. Documenta Matematica, s://ui.adsabs.harvard.edu/abs/1986Natur.323..533R).
Extra Volume ISMP. pp. 389–400. S2CID 15568746 (https://api.sem doi:10.1038/323533a0 (https://doi.org/10.1038%2F323533a0).
anticscholar.org/CorpusID:15568746). S2CID 205001834 (https://api.semanticscholar.org/CorpusID:20500
1834).
8. Griewank, Andreas; Walther, Andrea (2008). Evaluating Derivatives:
Principles and Techniques of Algorithmic Differentiation, Second 22. Tan, Hong Hui; Lim, King Han (2019). "Review of second-order
Edition (https://books.google.com/books?id=xoiiLaRxcbEC). SIAM. optimization techniques in artificial neural networks
ISBN 978-0-89871-776-1. backpropagation" (https://doi.org/10.1088%2F1757-899X%2F495%
2F1%2F012003). IOP Conference Series: Materials Science and
9. Schmidhuber, Jürgen (2015). "Deep learning in neural networks: An
overview". Neural Networks. 61: 85–117. arXiv:1404.7828 (https://a Engineering. 495 (1): 012003. Bibcode:2019MS&E..495a2003T (htt
ps://ui.adsabs.harvard.edu/abs/2019MS&E..495a2003T).
rxiv.org/abs/1404.7828). doi:10.1016/j.neunet.2014.09.003 (https://d
doi:10.1088/1757-899X/495/1/012003 (https://doi.org/10.1088%2F1
oi.org/10.1016%2Fj.neunet.2014.09.003). PMID 25462637 (https://p
ubmed.ncbi.nlm.nih.gov/25462637). S2CID 11715509 (https://api.s 757-899X%2F495%2F1%2F012003). S2CID 208124487 (https://ap
emanticscholar.org/CorpusID:11715509). i.semanticscholar.org/CorpusID:208124487).
10. Schmidhuber, Jürgen (2015). "Deep Learning" (https://doi.org/10.42 23. Wiliamowski, Bogdan; Yu, Hao (June 2010). "Improved
49%2Fscholarpedia.32832). Scholarpedia. 10 (11): 32832. Computation for Levenberg–Marquardt Training" (https://www.eng.a
Bibcode:2015SchpJ..1032832S (https://ui.adsabs.harvard.edu/abs/ uburn.edu/~wilambm/pap/2010/Improved%20Computation%20for%
20LM%20Training.pdf) (PDF). IEEE Transactions on Neural
2015SchpJ..1032832S). doi:10.4249/scholarpedia.32832 (https://do
i.org/10.4249%2Fscholarpedia.32832). Networks and Learning Systems. 21 (6).
24. Martens, James (August 2020). "New Insights and Perspectives on
11. Goodfellow, Bengio & Courville (2016, p. 217 (https://www.deeplear
ningbook.org/contents/mlp.html#pf36)–218), "The back-propagation the Natural Gradient Method". Journal of Machine Learning
algorithm described here is only one approach to automatic Research (21). arXiv:1412.1193 (https://arxiv.org/abs/1412.1193).
differentiation. It is a special case of a broader class of techniques 25. Nielsen (2015), "[W]hat assumptions do we need to make about our
called reverse mode accumulation." cost function ... in order that backpropagation can be applied? The
first assumption we need is that the cost function can be written as
12. Rosenblatt, Frank (1962). Principles of Neurodynamics:
an average ... over cost functions ... for individual training examples
Perceptrons and the Theory of Brain Mechanisms Cornell
Aeronautical Laboratory. Report no. VG-1196-G-8 Report (Cornell ... The second assumption we make about the cost is that it can be
Aeronautical Laboratory). Spartan. pp. Page XIII Table of contents, written as a function of the outputs from the neural network ..."
Page 292 "13.3 Back-Propagating Error Correction Procedures", 26. LeCun, Yann; Bengio, Yoshua; Hinton, Geoffrey (2015). "Deep
Page 301 "figure 39 BACK-PROPAGATING ERROR- learning". Nature. 521 (7553): 436–444.
CORRECTION EXPERIMENTS". Bibcode:2015Natur.521..436L (https://ui.adsabs.harvard.edu/abs/20
15Natur.521..436L). doi:10.1038/nature14539 (https://doi.org/10.10
13. Kelley, Henry J. (1960). "Gradient theory of optimal flight paths".
ARS Journal. 30 (10): 947–954. doi:10.2514/8.5282 (https://doi.org/ 38%2Fnature14539). PMID 26017442 (https://pubmed.ncbi.nlm.nih.
10.2514%2F8.5282). gov/26017442). S2CID 3074096 (https://api.semanticscholar.org/Co
rpusID:3074096).
14. Bryson, Arthur E. (1962). "A gradient method for optimizing multi-
27. Buckland, Matt; Collins, Mark (2002). AI Techniques for Game
stage allocation processes". Proceedings of the Harvard Univ.
Programming. Boston: Premier Press. ISBN 1-931841-08-X.
Symposium on digital computers and their applications, 3–6 April
1961. Cambridge: Harvard University Press. OCLC 498866871 (htt 28. Rodríguez, Omar Hernández; López Fernández, Jorge M. (2010).
ps://www.worldcat.org/oclc/498866871). "A Semiotic Reflection on the Didactics of the Chain Rule" (https://s
cholarworks.umt.edu/tme/vol7/iss2/10/). The Mathematics
15. Goodfellow, Bengio & Courville 2016, p. 214 (https://www.deeplear
Enthusiast. 7 (2): 321–332. doi:10.54870/1551-3440.1191 (https://d
ningbook.org/contents/mlp.html#pf33), "This table-filling strategy is
oi.org/10.54870%2F1551-3440.1191). S2CID 29739148 (https://ap
sometimes called dynamic programming."
i.semanticscholar.org/CorpusID:29739148). Retrieved 2019-08-04.
29. Rosenblatt, Frank (1962). Principles of Neurodynamics. Spartan,
New York.
30. Amari, Shun'ichi (1967). "A theory of adaptive pattern classifier". 38. Dreyfus, Stuart (1973). "The computational solution of optimal
IEEE Transactions. EC (16): 279–307. control problems with time lag". IEEE Transactions on Automatic
31. Werbos, Paul (1982). "Applications of advances in nonlinear Control. 18 (4): 383–385. doi:10.1109/tac.1973.1100330 (https://doi.
sensitivity analysis" (http://werbos.com/Neural/SensitivityIFIPSepte org/10.1109%2Ftac.1973.1100330).
mber1981.pdf) (PDF). System modeling and optimization. Springer. 39. Parker, D.B. (1985). "Learning Logic". Center for Computational
pp. 762–770. Archived (https://web.archive.org/web/201604140555 Research in Economics and Management Science. Cambridge MA:
03/http://werbos.com/Neural/SensitivityIFIPSeptember1981.pdf) Massachusetts Institute of Technology.
(PDF) from the original on 14 April 2016. Retrieved 2 July 2017. 40. Hertz, John (1991). Introduction to the theory of neural computation.
32. Werbos, Paul J. (1994). The Roots of Backpropagation : From Krogh, Anders., Palmer, Richard G. Redwood City, Calif.: Addison-
Ordered Derivatives to Neural Networks and Political Forecasting. Wesley. p. 8. ISBN 0-201-50395-6. OCLC 21522159 (https://www.w
New York: John Wiley & Sons. ISBN 0-471-59897-6. orldcat.org/oclc/21522159).
33. Rumelhart, David E.; Hinton, Geoffrey E.; Williams, Ronald J. 41. Wan, Eric A. (1994). "Time Series Prediction by Using a
(1986b). "8. Learning Internal Representations by Error Connectionist Network with Internal Delay Lines". In Weigend,
Propagation" (https://archive.org/details/paralleldistribu00rume). In Andreas S.; Gershenfeld, Neil A. (eds.). Time Series Prediction :
Rumelhart, David E.; McClelland, James L. (eds.). Parallel Forecasting the Future and Understanding the Past. Proceedings of
Distributed Processing : Explorations in the Microstructure of the NATO Advanced Research Workshop on Comparative Time
Cognition. Vol. 1 : Foundations. Cambridge: MIT Press. ISBN 0- Series Analysis. Vol. 15. Reading: Addison-Wesley. pp. 195–217.
262-18120-7. ISBN 0-201-62601-2. S2CID 12652643 (https://api.semanticschola
34. Alpaydin, Ethem (2010). Introduction to Machine Learning (https://b r.org/CorpusID:12652643).
ooks.google.com/books?id=4j9GAQAAIAAJ). MIT Press. 42. Chang, Franklin; Dell, Gary S.; Bock, Kathryn (2006). "Becoming
ISBN 978-0-262-01243-0. syntactic". Psychological Review. 113 (2): 234–272.
35. Dreyfus, Stuart (1962). "The numerical solution of variational doi:10.1037/0033-295x.113.2.234 (https://doi.org/10.1037%2F0033
problems" (https://doi.org/10.1016%2F0022-247x%2862%2990004 -295x.113.2.234). PMID 16637761 (https://pubmed.ncbi.nlm.nih.go
-5). Journal of Mathematical Analysis and Applications. 5 (1): 30– v/16637761).
45. doi:10.1016/0022-247x(62)90004-5 (https://doi.org/10.1016%2F 43. Janciauskas, Marius; Chang, Franklin (2018). "Input and Age-
0022-247x%2862%2990004-5). Dependent Variation in Second Language Learning: A
36. Dreyfus, Stuart E. (1990). "Artificial Neural Networks, Back Connectionist Account" (https://www.ncbi.nlm.nih.gov/pmc/articles/
Propagation, and the Kelley-Bryson Gradient Procedure". Journal PMC6001481). Cognitive Science. 42 (Suppl Suppl 2): 519–554.
of Guidance, Control, and Dynamics. 13 (5): 926–928. doi:10.1111/cogs.12519 (https://doi.org/10.1111%2Fcogs.12519).
Bibcode:1990JGCD...13..926D (https://ui.adsabs.harvard.edu/abs/1 PMC 6001481 (https://www.ncbi.nlm.nih.gov/pmc/articles/PMC6001
990JGCD...13..926D). doi:10.2514/3.25422 (https://doi.org/10.251 481). PMID 28744901 (https://pubmed.ncbi.nlm.nih.gov/28744901).
4%2F3.25422). 44. Fitz, Hartmut; Chang, Franklin (2019). "Language ERPs reflect
37. Mizutani, Eiji; Dreyfus, Stuart; Nishio, Kenichi (July 2000). "On learning through prediction error propagation". Cognitive
derivation of MLP backpropagation from the Kelley-Bryson optimal- Psychology. 111: 15–52. doi:10.1016/j.cogpsych.2019.03.002 (http
control gradient formula and its application" (https://coeieor.wpengin s://doi.org/10.1016%2Fj.cogpsych.2019.03.002).
e.com/wp-content/uploads/2019/03/ijcnn2k.pdf) (PDF). hdl:21.11116/0000-0003-474D-8 (https://hdl.handle.net/21.11116%
Proceedings of the IEEE International Joint Conference on Neural 2F0000-0003-474D-8). PMID 30921626 (https://pubmed.ncbi.nlm.ni
Networks. h.gov/30921626). S2CID 85501792 (https://api.semanticscholar.or
g/CorpusID:85501792).
45. "Photonic Chips Curb AI Training's Energy Appetite - IEEE
Spectrum" (https://spectrum.ieee.org/backpropagation-optical-ai).
spectrum.ieee.org. Retrieved 2023-05-25.

Further reading
Goodfellow, Ian; Bengio, Yoshua; Courville, Aaron (2016). "6.5 Back-Propagation and Other Differentiation Algorithms" (https://www.deeplearn
ingbook.org/contents/mlp.html#pf25). Deep Learning (http://www.deeplearningbook.org). MIT Press. pp. 200–220. ISBN 9780262035613.
Nielsen, Michael A. (2015). "How the backpropagation algorithm works" (http://neuralnetworksanddeeplearning.com/chap2.html). Neural
Networks and Deep Learning (http://neuralnetworksanddeeplearning.com). Determination Press.
McCaffrey, James (October 2012). "Neural Network Back-Propagation for Programmers" (https://docs.microsoft.com/en-us/archive/msdn-mag
azine/2012/october/test-run-neural-network-back-propagation-for-programmers). MSDN Magazine.
Rojas, Raúl (1996). "The Backpropagation Algorithm" (https://page.mi.fu-berlin.de/rojas/neural/chapter/K7.pdf) (PDF). Neural Networks : A
Systematic Introduction. Berlin: Springer. ISBN 3-540-60505-3.

External links
Backpropagation neural network tutorial at the Wikiversity
Bernacki, Mariusz; Włodarczyk, Przemysław (2004). "Principles of training multi-layer neural network using backpropagation" (http://galaxy.ag
h.edu.pl/~vlsi/AI/backp_t_en/backprop.html).
Karpathy, Andrej (2016). "Lecture 4: Backpropagation, Neural Networks 1" (https://www.youtube.com/watch?v=i94OvYb6noo&list=PLkt2uSq6
rBVctENoVBg1TpCC7OQi31AlC&index=4). CS231n. Stanford University. Archived (https://ghostarchive.org/varchive/youtube/20211212/i94
OvYb6noo) from the original on 2021-12-12 – via YouTube.
"What is Backpropagation Really Doing?" (https://www.youtube.com/watch?v=Ilg3gGewQ5U&list=PLZHQObOWTQDNU6R1_67000Dx_ZCJ
B-3pi&index=3). 3Blue1Brown. November 3, 2017. Archived (https://ghostarchive.org/varchive/youtube/20211212/Ilg3gGewQ5U) from the
original on 2021-12-12 – via YouTube.
Putta, Sudeep Raja (2022). "Yet Another Derivation of Backpropagation in Matrix Form" (https://sudeepraja.github.io/BackpropAdjoints/).

Retrieved from "https://en.wikipedia.org/w/index.php?title=Backpropagation&oldid=1165816909"