In more practical terms neural networks are Non-linear Statistical Data Modeling tools. They can be used to model complex relationships between inputs and outputs or to Find Patterns in data.
s in the Human Brain .]]
.]]
There is no precise agreed definition among researchers as to what a Neural Network is, but most would agree that it involves a network of simple processing elements ( Neurons ) which can exhibit complex global behaviour, determined by the connections between the processing elements and element parameters. The original inspiration for the technique was from examination of the Central Nervous System and the neurons (and their Axons , Dendrites and Synapses ) which constitute one of its most significant information processing elements (see Neuroscience ). In a neural network model, simple Nodes (called variously "neurons", "neurodes", "PEs" ("processing elements") or "units") are connected together to form a network of nodes — hence the term "neural network". While a neural network does not have to be adaptive per se, its practical use comes with algorithms designed to alter the strength (weights) of the connections in the network to produce a desired signal flow.
These networks are also similar to the Biological Neural Networks in the sense that functions are performed collectively and in parallel by the units, rather than there being a clear delineation of sub-tasks to which various units are assigned (see also Connectionism ). Currently, the term ANN tends to refer mostly to neural network models employed in Statistics and Artificial Intelligence . Neural Network models designed with emulation of the Central Nervous System (CNS) in mind are a subject of Theoretical Neuroscience .
In modern Software Implementations of artificial neural networks the approach inspired by biology has more or less been abandoned for a more practical approach based on statistics and signal processing. In some of these systems neural networks, or parts of neural networks (such as Artificial Neuron s) are used as components in larger systems that combine both adaptive and non-adaptive elements. While the more general approach of such Adaptive Systems is more suitable for real-world problem solving, it has far less to do with the traditional artificial intelligence connectionist models. What they do however have in common is the principle of non-linear, distributed, parallel and local processing and adaptation.
Neural network models in artificial intelligence are usually referred to as artificial neural networks (ANNs); these are essentially simple mathematical models defining a function . Each type of ANN model corresponds to a ''class'' of such functions.
The word ''network'' in the term 'artificial neural network' arises because the function is defined as a composition of other functions , which can further be defined as a composition of other functions. This can be conveniently represented as a network structure, with arrows depicting the dependencies between variables. A widely used type of composition is the ''nonlinear weighted sum'', where , where is some predefined function, such as the Hyperbolic Tangent . It will be convenient for the following to refer to a collection of functions as simply a vector .
This figure depicts such a decomposition of , with dependencies between variables indicated by arrows. These can be interpreted in two ways.
The first view is the functional view: the input is transformed into a 3-dimensional vector , which is then transformed into a 2-dimensional vector , which is finally transformed into . This view is most commonly encountered in the context of Optimization .
The second view is the probabilistic view: the Random Variable depends upon the random variable , which depends upon , which depends upon the random variable . This view is most commonly encountered in the context of Graphical Models .
The two views are largely equivalent. In either case, for this particular network architecture, the components of individual layers, i.e. the components of are independent of each other given their input . This naturally enables a degree of parallelism in the implementation.
Networks such as the previous one are commonly called Feedforward , because their graph is a Directed Acyclic Graph . Networks with cycles are commonly called Recurrent . Such networks are commonly depicted in the manner shown at the top of the figure above, where is shown as being dependent upon itself. However, there is an implied temporal dependence which is not shown. What this actually means in practice is that the value of at some point in time depends upon the values of at one or more other points in time. The graphical model at the bottom of the figure illustrates the case the value of at time only depends upon its last value. Models such as these, which have no dependencies in the future, are called Causal Models .
See Also: graphical models
However interesting such functions may be in themselves, what has attracted the most interest in neural networks is the possibility of ''learning'', which in practice means the following:
- \in F which solves the task in an ''optimal sense''.
- ,
The Cost Function is an important concept in learning, as it is a measure of how far away we are from an optimal solution to the problem that we want to solve. Learning algorithms search through the solution space in order to find a function that has the smallest possible cost.
|
| |
\exp(w_0\x-w\^2)</math>, with <math>w, x \in \mathbb{R}^n</math><math>w_0 \in \mathbb{R}</math>
|
While it is possible to arbitrarily define some
Ad Hoc cost function, frequently a particular cost will be used either because it has desirable properties (such as convexity) or because it arises naturally from a particular formulation of the problem (i.e. in a probabilistic formulation the posterior probability of the model can be used as an inverse cost). Ultimately, the cost function will depend on the task we wish to perform. The three main categories of learning tasks are overviewed below.
There are three major learning paradigms, each corresponding to a particular abstract learning task. These are
Supervised Learning ,
Unsupervised Learning and
Reinforcement Learning . Usually any given type of network architecture can be employed in any of those tasks.
In
Supervised Learning , we are given a set of example pairs
and the aim is to find a function f in the allowed class of functions that matches the examples. In other words, we wish to ''infer'' the mapping implied by the data; the cost function is related to the mismatch between our mapping and the data and it implicitly contains prior knowledge about the problem domain.
A commonly used cost is the
Mean-squared Error which tries to minimise the average error between the network's output, f(x), and the target value y over all the example pairs. When one tries to minimise this cost using
Gradient Descent for the class of neural networks called Multi-Layer Perceptrons, one obtains the well-known
Backpropagation Algorithm for training neural networks.
Tasks that fall within the paradigm of supervised learning are
Pattern Recognition (also known as classification) and
Regression (also known as function approximation). The supervised learning paradigm is also applicable to sequential data, i.e. for speech and for gesture recognition.
In
Unsupervised Learning we are given some data
, and the cost function to be minimised can be any function of the data
and the network's output,
.
The cost function is dependent on the task (what we are trying to model) and our ''a priori'' assumptions (the implicit properties of our model, its parameters and the observed variables).
As a trivial example, consider the model
, where
is a constant and the cost
. Minimising this cost will give us a value of
that is equal to the mean of the data. The cost function can be much more complicated. Its form depends on the application: For example in compression it could be related to the
Mutual Information between x and y. In statistical modelling, it could be related to the
Posterior Probability of the model given the data. (Note that in both of those examples those quantities would be maximised rather than minimised)
Tasks that fall within the paradigm of unsupervised learning are in general
Estimation problems; the applications include
Clustering , the estimation of
Statistical Distributions ,
Compression and
Filtering .
In
Reinforcement Learning , data
is usually not given, but generated by an agent's interactions with the environment. At each point in time
, the agent performs an action
and the environment generates an observation
and an instantaneous cost
, according to some (usually unknown) dynamics. The aim is to discover a ''policy'' for selecting actions that minimises some measure of a long-term cost, i.e. the expected cumulative cost. The environment's dynamics and the long-term cost for each policy are usually unknown, but can be estimated.
More formally, the environment is defined as a (MC). The aim is to discover the policy that minimises the cost, i.e. the MC for which the cost is minimal.
ANNs are frequently used in reinforcement learning as part of the overall algorithm.
Tasks that fall within the paradigm of reinforcement learning are
Control problems,
Games and other
Sequential Decision Making tasks.
See also:
Dynamic Programming ,
Stochastic Control
There are numerous algorithms available for training neural network models; most of them can be viewed as a straightforward application of
Optimization theory and
Statistical Estimation .
Most of the algorithms used in training artificial neural networks are employing some form of
Gradient Descent . This is done by simply taking the derivative of the cost function with respect to the network parameters and then changing those parameters in a
Gradient-related direction.
Evolutionary Methods ,
Simulated Annealing , and
Expectation-maximization and
Non-parametric Methods are among other commonly used methods for training neural networks. See also
Machine Learning .
Perhaps the greatest advantage of ANNs is their ability to be used as an arbitrary function approximation mechanism which 'learns' from observed data. However, using them is not so straightforward and a relatively good understanding of the underlying theory is essential.
- Choice of model: This will depend on the data representation and the application. Overly complex models tend to lead to problems with learning.
- Learning algorithm: There are numerous tradeoffs between learning algorithms. Almost any algorithm will work well with the ''correct hyperparameters'' for training on a particular fixed dataset. However selecting and tuning an algorithm for training on unseen data requires a significant amount of experimentation.
- Robustness: If the model, cost function and learning algorithm are selected appropriately the resulting ANN can be extremely robust.
With the correct implementation ANNs can be used naturally in
Online Learning and large dataset applications. Their simple implementation and the existence of mostly local dependencies exhibited in the structure allows for fast, parallel implementations in hardware.
The utility of artificial neural network models lies in the fact that they can be used to infer a function from observations. This is particularly useful in applications where the complexity of the data or task makes the design of such a function by hand impractical.
The tasks to which artificial neural networks are applied tend to fall within the following broad categories:
Application areas include system identification and control (vehicle control, process control), game-playing and decision making (backgammon, chess, racing), pattern recognition (radar systems, face identification, object recognition and more), sequence recognition (gesture, speech, handwritten text recognition), medical diagnosis, financial applications,
Data Mining (or knowledge discovery in databases, "KDD"), visualisation and
E-mail Spam filtering.
''Main article:''
Neural Network Software
is used to
Simulate ,
Research ,
Develop and apply
Artificial Neural Network s,
Biological Neural Network s and in some cases a wider array of
Adaptive System s.
The
Feedforward Neural Networks are the first and arguably simplest type of artificial neural networks devised. In this network, the information moves in only one direction, forward, from the input nodes, through the hidden nodes (if any) and to the output nodes. There are no cycles or loops in the network.
The earliest kind of neural network is a ''single-layer perceptron'' network, which consists of a single layer of output nodes; the inputs are fed directly to the outputs via a series of weights. In this way it can be considered the simplest kind of feed-forward network. The sum of the products of the weights and the inputs is calculated in each node, and if the value is above some threshold (typically 0) the neuron fires and takes the activated value (typically 1); otherwise it takes the deactivated value (typically -1). Neurons with this kind of activation function are also called ''McCulloch-Pitts neurons'' or ''threshold neurons''. In the literature the term ''
Perceptron '' often refers to networks consisting of just one of these units. They were described by
Warren McCulloch and
Walter Pitts in the
1940s .
A perceptron can be created using any values for the activated and deactivated states as long as the threshold value lies between the two. Most perceptrons have outputs of 1 or -1 with a threshold of 0 and there is some evidence that such networks can be trained more quickly than networks created from nodes with different activation and deactivation values.
Perceptrons can be trained by a simple learning algorithm that is usually called the ''
Delta Rule ''. It calculates the errors between calculated output and sample output data, and uses this to create an adjustment to the weights, thus implementing a form of
Gradient Descent .
Single-unit perceptrons are only capable of learning linearly separable patterns; in
1969 in a famous
Monograph entitled ''
Perceptrons '' by
Marvin Minsky and
Seymour Papert showed that it was impossible for a single-layer perceptron network to learn an
XOR Function . They conjectured (incorrectly) that a similar result would hold for a multi-layer perceptron network. Although a single threshold unit is quite limited in its computational power, it has been shown that networks of parallel threshold units can approximate any continuous function from a compact interval of the real numbers into the interval
This very recent result can be found in [Auer, Burgsteiner, Maass: The p-delta learning rule for parallel perceptrons, 2001 (state Jan 2003: submitted for publication) .
A single-layer neural network can compute a continuous output instead of a
Step Function . A common choice is the so-called
Logistic Function :
:
With this choice, the single-layer network is identical to the
Logistic Regression model, widely used in
Statistical Modelling . The
Logistic Function is also known as the
Sigmoid Function . It has a continuous derivative, which allows it to be used in backpropagation.
This class of networks consists of multiple layers of computational units, usually interconnected in a feed-forward way. Each neuron in one layer has directed connections to the neurons of the subsequent layer. In many applications the units of these networks apply a sigmoid function as an activation function.
The ''universal approximation theorem'' for neural networks states that every continuous function that maps intervals of real numbers to some output interval of real numbers can be approximated arbitrarily closely by a multi-layer perceptron with just one hidden layer. This result holds only for restricted classes of activation functions, e.g. for the sigmoidal functions.
Multi-layer networks use a variety of learning techniques, the most popular being ''
Back-propagation ''. Here the output values are compared with the correct answer to compute the value of some predefined error-function. By various techniques the error is then fed back through the network. Using this information, the algorithm adjusts the weights of each connection in order to reduce the value of the error function by some small amount. After repeating this process for a sufficiently large number of training cycles the network will usually converge to some state where the error of the calculations is small. In this case one says that the network has ''learned'' a certain target function. To adjust weights properly one applies a general method for non-linear
Optimization task that is called
Gradient Descent . For this, the derivative of the error function with respect to the network weights is calculated and the weights are then changed such that the error decreases (thus going downhill on the surface of the error function). For this reason back-propagation can only be applied on networks with differentiable activation functions.
In general the problem of teaching a network that performs well, even on samples that were not used as training samples, is a quite subtle issue that requires additional techniques. This is especially important for cases where only very limited numbers of training samples are available. The danger is that the network
Overfits the training data and fails to capture the true statistical process generating the data.
Computational Learning Theory is concerned with training classifiers on a limited amount of data. In the context of neural networks a simple
Heuristic , called
Early Stopping , often ensures that the network will generalize well to examples not in the training set.
Other typical problems of the back-propagation algorithm are the speed of convergence and the possibility to end up in a
Local Minimum of the error function. Today there are practical solutions that make back-propagation in multi-layer perceptrons the solution of choice for many
Machine Learning tasks.
Adaptive Linear Neuron or later called Adaptive Linear Element. It was developed by Professor
Bernard Widrow and his graduate student
Ted Hoff at
Stanford University in 1960. It's based on McCulloch-Pitts model. It consists of a weight, a bias and a summation function.
Operation:
Its adaptation is defined through a cost function (error metric) of the residual
where
is the desired input. With the
MSE error metric
the adapted weight and bias become:
and
While the Adaline is through this capable of simple linear regression, it has limited practical use.
There is an extension of the Adaline, called the Multiple Adaline (MADALINE)that consists of two or more adalines serially connected.
See Also: Radial basis function
Radial Basis Functions are powerful techniques for interpolation in multidimensional space. A RBF is a function which has built into a distance criterion with respect to a centre. Radial basis functions have been applied in the area of neural networks where they may be used as a replacement for the sigmoidal hidden layer transfer function in multilayer perceptrons. RBF networks have 2 layers of processing: In the first, input is mapped onto each RBF in the 'hidden' layer. The RBF chosen is usually a Gaussian. In regression problems the output layer is then a linear combination of hidden layer values representing mean predicted output. The interpretation of this output layer value is the same as a regression model in statistics. In classification problems the output layer is typically a sigmoid function of a linear combination of hidden layer values, representing a posterior probability. Performance in both cases is often improved by shrinkage techniques, known as ridge regression in classical statistics and known to correspond to a prior belief in small parameter values (and therefore smooth output functions) in a Bayesian framework.
RBF networks have the advantage of not suffering from local minima in the same way as multilayer perceptrons. This is because the only parameters that are adjusted in the learning process are the linear mapping from hidden layer to output layer. Linearity ensures that the error surface is quadratic and therefore has a single easily found minimum. In regression problems this can be found in one matrix operation. In classification problems the fixed non-linearity introduced by the sigmoid output function is most efficiently dealt with using iterated reweighted least squares.
RBF networks have the disadvantage of requiring good coverage of the input space by radial basis functions. RBF centres are determined with reference to the distribution of the input data, but without reference to the prediction task. As a result, representational resources may be wasted on areas of the input space that are irrelevant to the learning task. A common solution is to associate each data point with its own centre, although this can make the linear system to be solved in the final layer rather large, and requires shrinkage techniques to avoid overfitting.
Associating each input datum with an RBF leads naturally to kernel methods such as
Support Vector Machines and Gaussian Processes (the RBF is the kernel function). All three approaches use a non-linear kernel function to project the input data into a space where the learning problem can be solved using a linear model. Like Gaussian Processes, and unlike SVMs, RBF networks are typically trained in a Maximum Likelihood framework by maximizing the probability (minimizing the error) of the data under the model. SVMs take a different approach to avoiding overfitting by avoiding maximizing instead a margin. RBF networks are outperformed in most classification applications by SVMs. In regression applications they can be competitive when the dimensionality of the input space is relatively small.
The ''
Self-organizing Map '' (SOM) invented by
Teuvo Kohonen uses a form of
Unsupervised Learning . A set of artificial neurons learn to map points in an input space to coordinates in an output space. The input space can have different dimensions and topology from the output space, and the SOM will attempt to preserve these.
Contrary to feedforward networks,
Recurrent Neural Network (RNs) are models with bi-directional data flow. While a feedforward network propagates data linearly from input to output, RNs also propagate data from later processing stages to earlier stages.
A ''simple recurrent network'' (SRN) is a variation on the multi-layer perceptron, sometimes called an "Elman network" due to its invention by
Jeff Elman . A three-layer network is used, with the addition of a set of "context units" in the input layer. There are connections from the middle (hidden) layer to these context units fixed with a weight of one. At each time step, the input is propagated in a standard feed-forward fashion, and then a learning rule (usually back-propagation) is applied. The fixed back connections result in the context units always maintaining a copy of the previous values of the hidden units (since they propagate over the connections before the learning rule is applied). Thus the network can maintain a sort of state, allowing it to perform such tasks as sequence-prediction that are beyond the power of a standard multi-layer perceptron.
In a ''fully recurrent network'', every neuron receives inputs from every other neuron in the network. These networks are not arranged in layers. Usually only a subset of the neurons receive external inputs in addition to the inputs from all the other neurons, and another disjunct subset of neurons report their output externally as well as sending it to all the neurons. These distinctive inputs and outputs perform the function of the input and output layers of a feed-forward or simple recurrent network, and also join all the other neurons in the recurrent processing.
The
Hopfield Network is a recurrent neural network in which all connections are symmetric. Invented by
John Hopfield in
1982 , this network guarantees that its dynamics will converge. If the connections are trained using
Hebbian Learning then the Hopfield network can perform robust content-addressable memory, robust to connection alteration.
A
Stochastic Neural Network differs from a regular neural network in the fact that it introduces random variations into the network. In a probabilistic view of neural networks, such random variations can be viewed as a form of
Statistical Sampling , such as
Monte Carlo Sampling .
The
Boltzmann Machine can be thought of as a noisy Hopfield network. Invented by
Geoff Hinton and
Terry Sejnowski in
1985 , the Boltzmann machine is important because it is one of the first neural networks to demonstrate learning of latent variables (hidden units). Boltzmann machine learning was slow to simulate, but the
Contrastive Divergence Algorithm of Geoff Hinton (circa
2000 ) allows models including Boltzmann machines and ''product of experts'' to be trained much faster.
Biological studies showed that the human brain functions not as one single massive network, but as a collection of small networks. This realisation gave birth to the concept of
Modular Neural Networks , in which several small networks cooperate or compete to solve problems.
A ''committee of machines'' (CoM) is a collection of different neural networks that together "vote" on a given example. This generally gives a much better result compared to other neural network models. In fact in many cases, starting with the same architecture and training but different initial random weights gives vastly different networks. A CoM tends to stabilize the result.
The CoM is similar to the general
Machine Learning ''
Bagging '' method, except that the necessary variety of machines in the committee is obtained by training from different random starting weights rather than training on different randomly selected subsets of the training data.
Is an extension of the ''committee of machines'' that goes beyond a simple/weighted average of different models.
ASNN represents a combination of an ensemble of feed-forward neural networks and the k-nearest neighbour technique (kNN). It uses the correlation between ensemble responses as a measure of distance amid the analysed cases for the kNN. This corrects the bias of the neural network ensemble. An associative neural network has a memory that can coincide with the training set. If new data becomes available, the network instantly improves its predictive ability and provides data approximation (self-learn the data) without a need to retrain the ensemble. Another important feature of ASNN is the possibility to interpret neural network results by analysis of correlations between data cases in the space of models. The method can be used on-line or downloaded at
www.vcclab.org .
These special networks do not fit in any of the previous categories.
''
Instantaneously Trained Neural Networks '' (ITNNs) are also called "Kak networks" after their inventor
Subhash Kak . They were inspired by the phenomenon of short-term learning that seems to occur instantaneously. In these networks the weights of the hidden and the output layers are mapped directly from the training vector data. Ordinarily, they work on binary data but versions for continuous data that require small additional processing are also available.
''Spiking (or pulsed) neural networks'' (SNNs) are models which explicitly take into account the timing of inputs. The network input and output are usually represented as series of spikes (delta function or more complex shapes). SNNs have an advantage of being able to continuously process information. They are often implemented as recurrent networks.
Networks of spiking neurons -- and the temporal correlations of neural assemblies in such networks -- have been used to model figure/ground separation and region linking in the visual system (see e.g. Reitboeck et.al.in Haken and Stadler: Synergetics of the Brain. Berlin, 1989).
Gerstner and Kistler have a freely-available online textbook on
Spiking Neuron Models .
Spiking neural networks with axonal conduction delays exhibit
Polychronization , and hence could have a potentially unlimited memory capacity.
In June 2005
IBM announced construction of a
Blue Gene Supercomputer dedicated to the simulation of a large recurrent spiking neural network
{Link without Title} .
Dynamic Neural Network s not only deal with nonlinear multivariate behaviour, but also include (learning of) time-dependent behaviour such as various transient phenomena and delay effects.
Meijer has a Ph.D. thesis online where regular feedforward perception networks are generalized with differential equations, using variable time step algorithms for learning in the time domain and including algorithms for learning in the frequency domain (in that case linearized around a set of static bias points).
These neural networks begin their training without any hidden neurons. As the output error reaches a predefined error threshold, the networks add a new hidden neuron. The new hidden neuron is connected to all input nodes, as well as, all previous hidden neurons. Training terminates when a suitable error threshold is reached or when the maximum number of hidden neurons is added.
A neuro-fuzzy networks is a
Fuzzy Inference System in the body of a
Artificial Neural Network . Depending on the ''FIS'' type, there are several layers that simulate the processes involved in a ''fuzzy inference'' like fuzzification, inference, aggregation and defuzzification. Embedding an ''FIS'' in a general structure of an ''ANN'' has the benefit of using available ''ANN'' training methods to find the parameters of a fuzzy system.
Artificial neural network models have a property called 'capacity', which roughly corresponds to their ability to model any given function. It is related to the amount of information that can be stored in the network and to the notion of complexity.
Nothing can be said in general about convergence, since it depends on a number of factors. Firstly, there may exist many local minima. This depends on the cost function and the model. Secondly, the optimisation method used might not be guaranteed to converge when far away from a local minimum. Thirdly, for a very large amount of data or parameters, some methods become impractical. In general, however, it has been found that theoretical guarantees regarding convergence are not always a very reliable guide to practical application.
In applications where the goal is to create a system that generalises well in unseen examples, the problem of overtraining has emerged. This arises in overcomplex or overspecified systems when the capacity of the network significantly exceeds the needed free parameters. There are two schools of thought for avoiding this problem: The first is to use cross-validation and similar techniques to check for the presence of overtraining and optimally select hyperparameters such as to minimise the generalisation error. The second is to use some form of ''regularisation''. This is a concept that emerges naturally in a probabilistic (Bayesian) framework, where the regularisation can be performed by putting a larger prior probability over simpler models; but also in statistical learning theory, where the goal is to minimise over two quantities: the 'empirical risk' and the 'structural risk', which roughly correspond to the error over the training set and the predicted error in unseen data due to overfitting.
Supervised neural networks that use an
MSE cost function can use formal statistical methods to determine the confidence of the trained model. The MSE on a validation set can be used as an estimate for variance. This value can then be used to calculate the
Confidence Interval of the output of the network, assuming a
Normal Distribution . A confidence analysis made this way is statistically valid as long as the output
Probability Distribution stays the same and the network is not modified.
By assigning a softmax activation function on the output layer of the neural network (or a softmax component in a component based neural network) for categorical target variables, the outputs can be interpreted as posterior probabilities. This is very useful in classification as it gives a certainty measure on classifications.
The softmax activation function:
Various techniques originally developed for studying disordered magnetic systems (
Spin Glasses ) have been successfully applied to simple neural network architectures, such as the perceptron. Influential work by E. Gardner and B. Derrida has revealed many interesting properties about perceptrons with real-valued synaptic weights, while later work by W. Krauth and M. Mezard has extended these principles to binary-valued synapses.
- Arima, et al., ,"''Neural network integrated circuit device having self-organizing function''". March 8, 1994.
- Bishop, C.M. (1995) ''Neural Networks for Pattern Recognition'', Oxford: Oxford University Press. ISBN 0-19-853849-9 (hardback) or ISBN 0-19-853864-2 (paperback)
- Duda, R.O., Hart, P.E., Stork, D.G. (2001) ''Pattern classification (2nd edition)'', Wiley, ISBN 0471056693
- Gurney, K. (1997) ''An Introduction to Neural Networks'' London: Routledge. ISBN 1-85728-673-1 (hardback) or ISBN 1-85728-503-4 (paperback)
- Haykin, S. (1999) '' Neural Networks: A Comprehensive Foundation'', Prentice Hall, ISBN 0-13-273350-1
- Hertz, J., Palmer, R.G., Krogh. A.S. (1990) ''Introduction to the theory of neural computation'', Perseus Books. ISBN 0201515601
- Lawrence, Jeanette (1994) ''Introduction to Neural Networks'', California Scientific Software Press. ISBN 1-883157-00-5
- Masters, Timothy (1994) ''Signal and Image Processing with Neural Networks'', John Wiley & Sons, Inc. ISBN 0-471-04963-8
- Smith, Murray (1993) ''Neural Networks for Statistical Modeling'', Van Nostrand Reinhold, ISBN 0-442-01310-8
- Wasserman, Philip (1993) ''Advanced Methods in Neural Computing'', Van Nostrand Reinhold, ISBN 0-442-00461-3
