Where are artificial neural networks applied?

The artificial neural networks have been applied to a number of problems in diverse fields such as engineering, finance, medical, physics, medicine, and biology and so on. 

- All these applications are based on the fact that these neural networks can simulate the human brain capabilities. 

- They have found a potential use in classification and prediction problems. 

- These networks can be classified under the non-linear data driven self adaptive approaches. 

- They come handy as a powerful tool when the underlying data relationship is not known. 

- They find it easy to recognize and learn the patterns and can correlate between the input sets and the result values.

- Once the artificial neural networks have been trained, they can be used in the prediction of the outcomes of the data. 

- They can even work when the data is not clear i.e., when it is noisy and imprecise. 

- This is the reason why they prove to be an ideal tool for modeling the agricultural data which is often very complex. 

- Their adaptive nature is their most important feature.

- It is because of this feature that the models developed using ANN is quite appealing when the data is available but there is a lack of understanding of the problem.

- These networks are particularly useful in those areas where the statistical methods can be employed. 

- They have uses in various fields:

    1. Classification Problems:

a)   Identification of underwater sonar currents.

b)   Speech recognition

c)   Prediction of the secondary structure of proteins.

d)   Remote sensing

e)   Image classification

f)    Speech synthesis

g)   ECG/ EMG/ EEG classification

h)   Data mining

i)     Information retrieval

j)    Credit card application screening

2. Time series applications:

a)   Prediction of stock market performance

b)   ARIMA time – series models

c)   Machine robot/ control manipulation

d)   Financial, engineering and scientific time series forecasting

e)   Inverse modeling of vocal tract

3. Statistical Applications:

a)   Discriminant analysis

b)   Logistic regression

c)   Bayes analysis

d)   Multiple regression

4. Optimization:

a)   Multiprocessor scheduling

b)   Task assignment

c)   VLSI routing

5. Real world Applications:

a)   Credit scoring

b)   Precision direct mailing

6. Business Applications:

a)   Real estate appraisal

b)   Credit scoring: It is used for determining the approval of a load as per the applicant’s information.

c)   Inputs

d)   Outputs

7. Mining Applications

a)   Geo-chemical modeling using neural pattern recognition technology.

8. Medical Applications:

a) Hospital patient stay length prediction system: the CRTS/ QURI system was developed using a neural network for predicting the number of days a patient has to stay in hospital. The major benefit of this system was that money was saved and better patient care. This system required the following 7 inputs:

Ø  Diagnosis

Ø  Complications and comorbidity

Ø  Body systems involved

Ø  Procedure codes and relationships

Ø  General health indicators

Ø  Patient demographics

Ø  Admission category

9. Management Applications: Jury summoning prediction: a system was developed that could predict the number of jurors that were actually required. Two inputs were supplied: the type of case and judge number. The system is known to have saved around 70 million.
10. Marketing Application: A neural network was developed for improving the direct mailing response rate. This network selected those individuals who were likely to respond to the 2^nd mailing. 9 variables were given as the input. It saved around 35 % of the total mailing cost.
11. Energy cost prediction: A neural network was developed that could predict the price of natural gas for the next month. It achieved an accuracy of 97%. 

What are advantages and disadvantages of artificial neural networks?

The artificial neural networks, since they can simulate the biological nervous system are used in many real life applications which are also their one of the biggest advantages. 

They have made it easy for carrying out complex processes such as:

Ø  Function approximation

Ø  Regression analysis

Ø  Time series prediction

Ø  Fitness approximation

Ø  Modeling

- With the artificial neuron networks, the classification based on sequence and pattern recognition along with other difficult things such as the sequential decision making and the novelty detection is possible. 

- A number of operations falling under the data processing category such as clustering, filtering, compression and blind source separation etc. are also carried out with the help of artificial neural networks. 

- Artificial neural networks can be considered as the backbone of the robotics engineering field. 

- They are used in computer numerical control and in directing the manipulators.

- It offers advantages in the following fields of:

1. System control (this including process control, vehicle control and natural resources management),
2. System identification,
3. Game – playing,
4. Quantum chemistry
5. Decision making (in games such as poker, chess, backgammon and so on.)
6. Pattern recognition (including face identification, radar systems, object recognition and so on.)
7. Sequence recognition (in handwritten text recognition, speech, gesture etc.)
8. Medical diagnosis
9. Financial applications i.e., in automated trading systems
10. Data mining
11. Visualization
12. E – mail spam filtering

- Today several types of cancers can be diagnosed using the artificial neural networks. 

- HLND is an ANN based hybrid system for detection of lung cancer. 

- The diagnosis carried out with this is more accurate plus the speed of radiology is more. 

- These diagnoses are then used for making some models based up on information of the patient. 

- Its following theoretical properties are nothing but an advantage to the industry:

1. Computational power: It provides a universal function approximator i.e., the multilayer perceptron or MLP.
2. Capacity: This property indicates about the ability of ANN to model almost any given function. It has a relation with both the notion of complexity and information contained in a network.
3. Convergence: This property is dependent on a number of factors such as:

Ø  Number of existing local minima which in turn depends up on model and the cost function.

Ø  Optimization method used

Ø  Impracticality of few methods for a large amount of parameters.

4. Generalization and statistics: Over training is quite a prominent problem in the applications where it is required to create a system that is capable of generalizing in unseen examples. This in turn leads to problem of the over specified or the convoluted systems along with the network exceeding the limit of the parameters. There are two solutions offered by ANN for this problem:

-   Croos – validation and

-   Regularization

Disadvantages of Artificial Neural Networks

1. It requires a lot of diverse training for making the artificial neural networks ready for the real world operations which is a drawback more prominent in the robotics industry.

2. Many storage and processing resources are required for implementing large software neural networks using ANNs.

3. The human has the ability to process the signals via a graph of neurons. A similar simulation of even a very small problem can call for excessive HD and RAM requirements.

4. Time and money cost for building ANNs is very large. 

5. Furthermore, simulation of the signal transmission through all the connections and associated neurons is required.     

What are types of artificial neural networks?

In this article we discuss the types of artificial neural networks. These models simulate the real life biological system of nervous system.

1. Feed forward neural network: 

- This is the simplest type of neural network that has been ever devised. 

- In these networks the information flow is unidirectional; therefore the data moves only in forward direction. 

- From input nodes data flows to the output nodes via hidden nodes (if there are any). 

- In this model there are no loops or cycles. 

- Different types of units can be used for constructing feed forward networks for example, McCulloch – pitts neurons.

- Continuous neurons are used in error back propagation along with the sigmoidal activation.

2. Radial basis function network: 

- For interpolating in a multi – dimensional space radial basis functions are the most powerful tools. 

- These functions can be built in to criterion of distance with respect to some center.

- These functions can be applied in the neural networks. 

- In these networks, sigmoidal hidden layer transfer characteristic can be replaced by these functions.

3. Kohonen self–organization network: 

- Un–supervised learning is performed with the help of self – organizing map or SOM. 

- This map was an invention of Teuvo Kohonen.

- Few neurons learn mapping points in the input space that could not coordinate in the output space. 

- The dimensions and topology of the input space can be different from those of the output space. SOM makes an attempt for preserving these.

4. Learning vector quantization or LVQ: 

- This can also be considered as neural network architecture. 

- This one also was a suggestion of Teuvo Kohonen.  

- In these prototypical representatives are parameterized along with two important things namely, a classification scheme based - up on distance and a distance measure.

5. Recurrent neural network: 

- These networks are somewhat contrary to the feed forward networks. 

- They offer a bi–directional flow of data.

- On a feed forward network data is propagated linearly from input to output. 

- Data from later stages of processing is also transferred to its earlier stages by this network. 

- Sometimes these also double up as the general sequence processors. 

- Recurrent neural networks have a number of types as mentioned below:

Ø  Fully recurrent network

Ø  Hopfield network

Ø  Boltzmann machine

Ø  Simple recurrent networks

Ø  Echo state network

Ø  Long short term memory network

Ø  Bi – directional RNN

Ø  Hierarchical RNN

Ø  Stochastic neural networks

6. Modular neural networks: 

- As per the studies have shown that human brain works actually as a collection of several small networks rather than as just one huge network, this ultimately helped in realizing the modular neural networks where smaller networks cooperate in solving a problem. 

- Modular networks are also of many types such as:

Ø  Committee of machines: Different networks that work together on a given problem are collectively termed as the committee of machines. The result achieved through this kind of networking is quite better than what is achieved with the others. The result is highly stabilized.

Ø  Associative neural network or ASNN: This is an extension of the previous one. And extends a little beyond the weighted average of various models. This one is a combined form of the k- nearest neighbor technique (kNN) and the feed forward neural networks. Its memory is coincident with that of the training set.

7. Physical neural network: 

- It consists of some resistance material that is electrically adjustable and capable of simulating the artificial synapses.

There are other types of ANNs that do not fall in any of the above categories:

Ø  Holographic associative memory

Ø  Instantaneously trained networks

Ø  Spiking neural networks

Ø  Dynamic neural networks

Ø  Cascading neural networks

Ø  Neuro – fuzzy networks

Ø  Compositional pattern producing networks

Ø  One – shot associative memory

What is an Artificial Neural Network (ANN)?

- The artificial neural network or ANN (sometimes also called as just neural network) is a mathematical model that has got its inspiration from the biological neural networks. 

- This network is supposed to consist of several artificial neurons that are interconnected. 

- This model works with a connectionist approach for computing and thus processes information based up on this only. 

- In a number of cases, the neural network can act as an adaptive system that has the ability of making changes in its structure while it is in some learning phase. 

- These networks are particularly used in searching patterns in data and for modeling the complex relationships that exist between the outputs and inputs. 

- An analogy to artificial neural network is the neuron network of the human brain. 

- In an ANN, the artificial nodes are termed as the neurons or sometimes as neurodes or units or the ‘processing elements’. 

- They are interconnected in such a way that they resemble a biological neural network. 

- Till now, no formal definition has been given for the artificial neural networks. - These processing elements or the neurons show a complex global behavior. 

- The connections between the neurons and their parameters is what that determines this behavior.

- There are certain algorithms that are designed for altering the strength of these connections in order to produce the desired flow of the signal. 

- The ANN operates up on these algorithms. 

- As in biological neural networks, in ANN also functions are performed in parallel and collectively by the processing units.

- Here, there is no delineation of the tasks that might be assigned to different units. 

- These neural networks are employed in various fields such as:

1. Statistics
2. Cognitive psychology
3. Artificial intelligence

- There are other neural network models that emulate biological CNS and are part of the following:

1. Computational neuroscience
2. Theoretical neuroscience

- The modern software implementation of the ANNs prefers a more practical approach than biologically inspired approach. 

- This practical approach is based up on the signal processing and statistics. The former approach has been largely abandoned. 

- Many times parts of these neural networks serve as components for the other larger systems that are a combination of non – adaptive and adaptive elements.

- Even though a more practical approach for solving the real world problems is the latter one, the former has more to do with the connectionist models of the traditional artificial intelligence. 

- Well the common thing between them is the principle of distributed, non – linear, local and parallel processing and adaptation. 

- A paradigm shift was marked by the use of neural networks during the late eighties. 

- This shift was from the high level artificial intelligence (expert systems) to low level machine learning (dynamical system). 

- These models are very simple and define functions such as:

f: X à Y

- Three types of parameters are used for defining an artificial neural network:

a)   The interconnection pattern between neuron layers

b)   The learning process

c)   The activation function

- The second parameter updates the weights of the connections and the third one converts the weighted input in to output. 

- Learning is the thing that has attracted many towards it. 

- There are 3 major learning paradigms that are offered by ANN:

1. Supervised learning
2. Un – supervised learning
3. Reinforcement learning

- Training a network requires selecting from a set of models that would best minimize the cost.

- A number of algorithms are available for training purpose where gradient descent is employed by most of the algorithms.

- Other methods available are simulated annealing, evolutionary methods and so on.