Here are lists of questions that people frequently ask NeuralWare about model development, NeuralWorks Predict, NeuralSight, and Professional II/PLUS. Please review these lists to see if your questions are answered here. If not, for sales and license related questions please contact us at This email address is being protected from spambots. You need JavaScript enabled to view it.. For technical and user interface related questions, please contact us at This email address is being protected from spambots. You need JavaScript enabled to view it..

General Model Development Questions

Back to Top


NeuralWorks Predict Questions

Back to Top


NeuralSight Questions

Back to Top


Professional II/PLUS Questions

Back to Top


Why should I consider using neural networks?

Neural networks are particularly well-suited for modeling complex non-linear relationships which cannot be accurately modeled by traditional statistical methods. The fundamental theoretical justfication for neural network technology is the Universal Approximation Theorem. In essence, the theorem states that any continuous function which maps a set of real numbers to another set of real numbers can be approximated to any required degree of accuracy by a feed-forward neural network with a single hidden layer and a finite number of hidden units which contain non-linear (e.g., sigmoid) transfer functions.

Back to Top


What kinds of applications are appropriate for neural networks?

Neural networks are empirical mathematical models, so they require data in order to learn patterns that are the basis for predictions, classifications, or clustering associations.  This means that basically any application domain that generates data is a candidate for neural network modeling. Some representative application domains include:

  • Financial markets (e.g., predicting security prices, interest rates, loan defaults).
  • Energy (predicting electricity demand, predicting best locations for drilling, classifying grid faults and other operational diagnostics)
  • Manufacturing (quality control, inventory forecasting)
  • Signal Processing (classifying radar signatures)
  • Medicine and Health Care (disease diagnosis, outcomes analysis, drug discovery)
  • Marketing (predicting response rates, customer preferences)
  • Process Control (predicting process states, materials classification)

If your modeling requirements can be addressed by classifications based on simple threshold values, or predictions based on linear regression models, then a neural network is probably not required. However, if you are faced with modeling more complex non-linear relationships, please contact NeuralWare to discuss your application domain and how NeuralWare technology can help!

Back to Top


How much data do I need to train a neural network?

The amount of data needed to train a neural network is very much problem-dependent. The quality of training data (i.e., how well the available training data represents the problem space) is as important as the quantity (i.e., the number of records, or examples of input-output pairs). The key is to use training data that generally span the problem data space. For relatively small datasets (fewer than 20 input variables, 100 to several thousand records) a minimum of 10 to 40 records (examples) per input variable is recommended for training. Note that if you are using products based on the NeuralWorks Predict Engine, the number of neural network inputs will likely be different from the number of input variables in the training dataset, due to data transformations that the Predict Engine may apply. When using NeuralWare products based on the Predict Engine, higher numbers of input records per field (e.g., 30 to 40) would be appropriate to account for data transformations. For relatively large datasets (more than 20 000 records), the dataset should be sub-sampled to obtain a smaller dataset that contains 30 - 50 records per input variable. In either case, any "extra" records should be used for validating the neural networks produced.

Back to Top


What is the difference between Training data, Testing data, Modeling data, and Validation data?

In NeuralWare terminology, Training data is data that is used to adjust the weights in a neural network. For a feed-forward neural network, the error that results when a Training data record is processed by the neural network is back-propagated through the network and used to adjust network weights. For a Self-Organizing Map, the "distance" between a Training data record and the weights of each processing element in the SOM is calculated and the shortest distance identifies the "winning" processing element. On the other hand, Testing data is data that is periodically used during the feed-forward network training process to determine whether training should terminate (NeuralWorks Professional II/PLUS and NeuralWorks Predict Engine), or addtional hidden units should be added and training should continue (NeuralWorks Predict Engine). Testing data is not used during the Self-Organizing Map training process since there is no "error" with respect to distance calculations.

Validation data is data that is held out and not used at all in the training process, either for feed-forward networks or Self-Organizing Maps. (Note that in the case of NeuralWorks Predict, if no data are explicity held out, then when training concludes the neural network is run with all available data to compute Validation performance metrics or to make a final assignment of SOM training records to SOM processing elements).

In NeuralWare documentation, the term Modeling data comprises Training data and Testing data. When you use products based on the Predict Engine to build neural networks, you specify how much of the available data should be used as Modeling data. If there is any remaining data, it will be treated as Validation data; or, if you specify that all available data should be used as Modeling data, then all data will also be used as Validation data. In either case, the Predict Engine will then further partition Modeling data into Training data and Testing data (by default, 70% of Modeling data is used for Training data, and 30% is used for Testing data). If you are using Professional II/PLUS, you explicitly specify how available data should be partitioned.

After a neural network is trained, it is run with Validation data in order to compute performance metrics. Running a neural network with Validation data provides the best indication of how well the network will perform when it is placed in service.

Back to Top


What is over-fitting, and why is it an issue?

Neural network learning is in essence fitting a curve (a complex mathematical function) to data. If the curve is made to match the Training data too precisely, then when other data (data that was not used for training, but still comes from the same problem space) are presented to the network, the results may not be acceptable. The reason for this is that rather than learning to generalize patterns in the training data, the network has essentially evolved into a look-up table (in other words, the neural network is too complex). It has "memorized" the input-output pairs used in training, but does not "know" how to process new inputs that are not in the look-up table - so the output is not an appropriate mapping for the inputs.

When you create neural networks with NeuralWorks Predict, available data is partitioned into Training data and Testing data. Periodically during the training process the current network is run with Testing data. If performance on Testing data is worse than a previous run with Testing data, Predict reverts to the previous network state and training terminates. At the conclusion of training, a quick test to determine whether the neural network has over-fit the Training data is to view the final architecture. If a network has more hidden units than it has inputs, there is a high probability that the network has over-fit the Training data.

When you create neural networks with Professional II/PLUS, the Savebest function is used to interrupt training and determine the network's performance on Testing data. As long as performance on the Testing data is improving, Professional II/PLUS saves the network on disc. When you set up the Savebest function for network training, you specify control parameters such as the total number of learning iterations to execute, the error convergence tolerance, and how many test functions to allow that do not improve performance before terminating training.

Back to Top


How do I choose the best neural network for my application?

There are several factors to consider when choosing the best neural network, and usually there are application-specific trade-offs which must be evaluated for their impact on the final system.

Generally the first factor to consider is a network's performance as measured by an appropriate metric and compared with some benchmark. For prediction networks, the metric might be Root Mean Square (RMS) Error or R Correlation, and the benchmark might be a regression model built from the same Modeling data. For classification networks, the metric might be Average Classification Accuracy, and the most basic benchmark would be a "flip of the coin" classification (i.e., 50% accuracy). You may also have other benchmarks, perhaps based on other modeling approaches, which can serve as baselines for measuring performance. After you have selected the metric (or metrics) and a benchmark value, you can then compare metrics computed for different networks to decide which is the better network.

The second factor to consider is the network architecture. The more compact a network is, the better it will generalize, as long as there are more neural network inputs than there are hidden units. In this context, it is important to recognize that the Predict Engine can produce neural networks which have no hidden units. If a network has no hidden units, it means that the Engine has identified data transformations that essentially capture the non-linear relationships in the data, and thus non-linear hidden units are not necessary.

A third factor to consider is the relative influence of input variables (i.e., the raw values of input fields in the training dataset) on model performance. For example, caution is warranted if a Sensitivity Analysis suggests that an input variable whose value is very volatile (e.g., because it is obtained from a sensor that is not very accurate) is also very influential with respect to a given neural network's output. In this case a different neural network which, although it exhibits slightly lower performance when performance metrics are compared, provides more stable output over-all and thus is a better choice.

A related factor is the ease with which input data for the selected neural network can be acquired. If the network with the best performance metric values also relies on input variables whose values are difficult to obtain, a different model that does not perform quite as well, but which also does not rely on inputs which are difficult to obtain, may be the better model.

Back to Top


How do I deploy my neural network application for use by others?

NeuralWare offers a variety of ways to deploy neural networks developed on NeuralWare platforms.

For neural networks created by the Predict Engine, deployment options include running the network in the Microsoft Excel environment with data in a spreadsheet, running the network in NeuralSight for batch processing of data in a file, running the network in a custom application that is linked with the Predict Run-Time Kit and an appropriate data source, or compiling a FlashCode representation of the network and linking it with a custom VisualBasic, C, or Fortran 77 application.

Linking a custom application to the RTK is by far the most flexible deployment option. When an application links to the RTK, if the neural network ever needs updating, the application simply has to load the updated network - there is no need to recompile and relink the application.

For neural networks created by Professional II/PLUS, deployment options include running the network in Professional II/PLUS for batch processing of data in a file, compiling a FlashCode representation of the network and linking it with a custom C application, or using the DesignerPack add-on for Professional II/PLUS to create a custom application that can both execute the network as well as retrain it.

Back to Top


When should I retrain my neural network?

After a neural network is deployed, it is important to monitor its performance to ensure that the network is still performing as expected. This means there must be an independent way to periodically compare the network's output (typically a prediction or classification result) with actual conditions and results in the problem domain. Also, it is important to archive this information, since it will be needed to create new modelng datasets if retraining is necessary. While this additional effort has costs, it is important to know if network performance has degraded and thus should be retrained. When performance degrades, either conditions in the problem domain have changed over time, or there may be an issue with the processes and procedures that are generating data for the neural network. At the same time, in many environments the original modeling data will continue to reflect the current operating conditions and retraining will not be necessary.

Back to Top


Which NeuralWare product should I purchase?

All NeuralWare development platforms generate robust neural networks. Generally the most important consideration is the nature of the modeling problem that you want to address.

If you are primarily interested in addressing a relatively specific modeling problem as quickly as possible, both in terms of developing models as well as in production deployment, then generally the NeuralWare platforms based on the NeuralWorks Predict Engine are optimal. The Predict Engine supports two neural network paradigms (feed-forward, for prediction and classification, and Self-Organizing Maps for clustering) which are appropriate for any modeling application. Furthermore, the Predict Engine's automated data preparation facilities eliminate the time consuming and error-prone processes of manually exploring the effects of data transformations and selecting input variables. For one-off or low-volume prototype modeling, with the option to deploy results in an Excel spreadsheet, NeuralWorks Predict with its Excel add-in offers a cost effective yet powerful development framework. For high-throughput comprehensive model and data exploration that involves large datasets and potentially thousands of models, NeuralSight is the platform of choice. It can shave weeks or months off the time needed to develop, evaluate, and deploy neural network models. While the up-front cost of NeuralSight is higher, your investment can be quickly returned in terms of the reduced time that senior scientists, analysts, or engineers need to spend on trial-and-error explorations of the impact that changes to data distributions or model parameters have on neural network performance.

In contrast, if your interest is more towards exploring the underlying mathematics and structure of neural networks, including paradigms other than feed-forward networks and Self-Organizing Maps, then the Professional II/PLUS platform is appropriate. Professional II/PLUS is particularly well suited for teaching and research in academic environments, or for modeling challenges for which feed-forward neural network or Self-Organizing Map performance has not been satisfactory. Professional II/PLUS provides low-level access to virtually all aspects of the architecture and learning strategies for each of the paradigms supported. As a consequence, its comprehensive user interface also requires somewhat more time to master, and optimization of network performance requires more direct involvement of the modeler. The capabilities of Professional II/PLUS are further enhanced by its two add-on packages, though they also require additional time and effort to effectively master. Designer Pack allows neural networks created in Professional II/PLUS to be packaged for deployment in customized C language applications that are externally compiled and linked. User-Defined Neuro-Dynamics supports creating new architectures and learning strategies which can then be re-integrated with Professional II/PLUS and used in other neural network development.

Also, note that NeuralWorks Predict and NeuralWorks Professional II/PLUS are often used together to explore the effectiveness of paradigms available only in Professional II/PLUS. The powerful and easy-to-use data pre-processing capabilities of Predict can be employed to generate a dataset that contains pre-processed and appropriately scaled values that then serve directly as inputs for neural networks supported by Professional II/PLUS. You can focus on optimizing neural network performance without needing to manually prepare the data.

Back to Top


When should I request 64-bit products?

The primary consideration related to choosing 64-bit products (Predict and NeuralSight) is whether you have a 64-bit version of Microsoft Excel installed (Excel 2010 or Excel 2013, installed in 64-bit mode) on a Windows 64-bit Operating System.

If your version of Excel is earlier than Excel (Office) 2010, it is a 32-bit version.

To determine whether Excel 2010 is 32-bit or 64-bit, open Excel, click the File menu, then select Help. On the right side of the page you will see information about Excel. If the description contains "64-bit", then Excel 2010 is installed in 64-bit mode. Otherwise it is installed in 32-bit model. The example image illustrates 64-bit mode.

To determine whether Excel 2013 is 32-bit or 64-bit, open Excel, click the File menu, then select Account. On the right side of the page, click About Excel. If the description contains "64-bit", then Excel 2013 is installed in 64-bit mode. Otherwise it installed in 32-bit mode. The example image illustrates 32-bit mode.

If you have a 64-bit version of Excel installed, then you should order 64-bit Predict (and NeuralSight, if applicable). Training time is typically reduced by 10% - 15% (64-bit versions are faster). In addition, if you do not have Excel, but you do use a 64-bit Windows Operating System, you can order 64-bit Predict (and 64-bit NeuralSight if applicable). Note that in this scenario you will only be able to use the Command Line Interface for Predict.

Otherwise (i.e., you have 32-bit Excel installed on either a 32-bit or a 64-bit Windows Operating System), you should order the 32-bit versions of Predict (and NeuralSight if applicable).

Back to Top


Can NeuralWare products be run on a server for multiple simultaneous users?

NeuralWare products are not designed as multi-user client/server applications. Product user interfaces generally appear on the monitor that is locally (directly) connected to the computer which has the running product installed. However, it is possible to run a NeuralWare product remotely using software such as Microsoft's remote desktop, or the Citrix GoToMyPC product.

This mode of use is permitted for single-user exclusive access under the terms of NeuralWare's standard license. In other words, only one person at a time may use a NeuralWare product, regardless of whether the person sitting directly at the computer which has the product installed, or the person is accessing the product remotely using third-party software. Note, however, that there are no built-in safeguards to prevent a remote user from over-writing or changing a file that a local user may also be accessing with another (non-NeuralWare) application.

Back to Top


How are NeuralWare products licensed?

NeuralWare base products (NeuralWorks Predict and NeuralWorks Professional II/PLUS) are licensed initially for use on a single computer (a single "seat") under a standard license which you must accept before using the products. After a neural network model has been developed and validated, NeuralWare offers a variety of negotiable license extensions that permit deploying the neural network on a computer other than the computer which has the base product installed. NeuralWare also offers significantly discounted licenses for degree-granting academic institutions. Licenses for NeuralPower are customized for each customer and executed by NeuralWare and the customer.

NeuralWare also offers a variety of licensing and purchase options for acredited degree-granting academic institutions.

Back to Top


NeuralWorks Predict Questions

What data limitations does Predict have?

The demonstration version of Predict limits datasets to 32 input fields and 512 data records. The demonstration version also prohibits saving networks, and does not offer the Command Line Interface (CLI)

Otherwise, the only inherent data limitation of the Predict Engine is that it permits a maximum of 4 000 input fields for a neural network.

When Predict is being used as an Excel add-in, data limitations are basically inherited from spreadsheet limitations. For versions of Excel earlier than Excel 2007, the number of input fields is limited to 255, and the number of data records is limited to 65 536 due to restrictions on the number of columns and rows permitted in a spreadsheet.

When Predict is being used as an Excel add-in with Excel 2007 or later, the number of input fields is limited to 4 000 (by the Predict Engine) and the number of data records is limited to 1 048 576 (the number of rows Excel permits in a spreadsheet).

The Predict Command Line Interface has the same limit of 4 000 input fields. The CLI enforces no limit on the number of data records that can be in a file; however, available computer memory and disc space may limit the size of files that Predict can process.

Back to Top


How does Predict handle missing or invalid data?

By default, when processing numeric fields NeuralWorks Predict substitutes the average value for missing or invalid field values (e.g., the field is empty, or contains alphanumeric characters). If the field is a symbolic field, Predict substitutes the symbolic value that is closest to the "average" symbolic value. This behavior can be modified, for example, to completely ignore records that have fields with missing or invalid data.

Back to Top


Does Predict require Microsoft Excel?

Predict does not explicitly require Microsoft Excel. If the Predict Installation Package does not find Microsoft Excel when Predict is installed, only the files which support the Predict Command Line Interface (CLI) are installed. The CLI provides access to the full range of features of the Predict Engine through a fully documented set of commands which you manually enter . The CLI can also process script files that you create.

However, if Excel is available on the computer where Predict is installed, the Excel Add-In Graphical User Interface offers much easier and more intuitive access to the power of the Predict Engine.

Back to Top


What types of neural networks can Predict produce?

The Predict Engine produces either feed-forward neural networks for prediction or classification types of problems, or Self-Organizing Maps for clustering types of problems.

If a feed-forward neural network is constructed, the Predict Engine automatically chooses the most appropriate internal objective function, and the network architecture is dynamically built using Cascade Correlation. The default objective function for prediction problems is Root Mean Square (RMS) Error. The default objective function for 2-class classification problems is a Bernoulli probability distribution. The default objective function for multi-class classification problems is Relative Entropy. While the type of network constructed for a specific problem type cannot be modified, the choice of internal objective function to use with a feed-forward network can be modified.

Back to Top


How long does it take to build a neural network with Predict?

The time required to build a neural network with Predict is very much dependent on the number of fields in a record and the number of records in the modeling dataset.

However, with current (2013) processor speeds in basic (mid-level) desktop or laptop computers running Windows 7 or 8, and accepting Predict's default modeling options, Predict will typically produce a network within at most several minutes. This assumes an Excel spreadsheet that contains up to several thousand records, with no more than 20 fields in each record.

In other words, you can very quickly create a single neural network that usually has good performance, in under 10 minutes.

Back to Top


How can I deploy a neural network created with Predict?

For neural networks created by the NeuralWorks Predict Engine, deployment options include running the network in the Microsoft Excel environment with data in a spreadsheet, running the network using the Command Line Interface  and data from a file, running the network in a custom application that is linked with the Predict Run-Time Kit and an appropriate data source, or compiling a FlashCode representation of the network and linking it with a custom VisualBasic, C, or Fortran 77 application.

If you want to deploy the network for use in an Excel spreadsheet that will be used on a computer other than the computer that created the network, a Run-Time Kit license and software are required for every computer that needs to run the network (in addition to having Excel on the computers).

If you want to deploy the network on another computer through an application that was created using a FlashCode representation of the network, a deployment license is required for every computer that needs to run the network. No additional software is required.

Back to Top


NeuralSight Questions

What data limitations does NeuralSight have?

NeuralSight has the same inherent data limitation of the Predict Engine regarding input fields in a dataset. The dataset can have a maximum of 4 000 input fields.

NeuralSight does not explicitly limit the number of records permitted in a dataset; however, seldom would there be a need for more than 10 000 records in a modeling dataset (see How much data do I need to train a neural network?). When a network is run, many more records may be processed without difficulty.

The size of Self-Organizing Maps that NeuralSight can process is limited by available memory. If your computer has at least 4 GB of memory, NeuralSight typically can process 30 x 30 SOMs with 25 - 40 input fields (you may need to increase the amount of memory allocated to Java in order to process large SOM models - how to do this is documented in the NeuralSight User Guide).

Back to Top


How does NeuralSight handle missing or invalid data?

NeuralSight data requirements are more stringent than the requirements imposed by the Predict Engine. If a record in a dataset contains fields with missing or invalid values, the record is not used for modeling or neural network evaluation. NeuralSight will display an error message when it encounters a record containing invalid or missing data, and the record itself is written to an error file so that you can view the record and determine what caused the problem.

Back to Top


Does NeuralSight require Microsoft Excel?

NeuralSight does NOT require Microsoft Excel. NeuralSight has a Java-based Graphical User Interface (GUI) that offers robust data and model performance visualization. At the same time, since NeuralSight does require NeuralWorks Predict, it is usually advantageous to have Excel also installed so that data and individual models can used in spreadsheets for quick assessments of performance.

Back to Top


What types of neural networks can NeuralSight produce?

NeuralSight is built on the Predict Engine, so it produces either feed-forward neural networks for prediction or classification problems, or Self-Organizing Maps for clustering problems. After you select the type of problem that you need a neural network for, NeuralSight will specify the problem to the Predict Engine, and all of the Predict Engine automation features will be employed.

If a feed-forward neural network is constructed, Predict automatically chooses the most appropriate internal objective function, and the network architecture is dynamically built using Cascade Correlation. The default objective function for prediction problems is Root Mean Square (RMS) Error. The default objective function for 2-class classification problems is a Bernoulli probability distribution. The default objective function for multi-class classification problems is Relative Entropy. While the type of network constructed for a specific problem type cannot be modified, the choice of internal objective function to use with a feed-forward network can be modified.

Back to Top


How long does it take to build a neural network with NeuralSight?

NeuralSight is designed to generate many neural network models without requiring you to specify the parameters and dataset partitioning for each neural network. For a reasonably-sized dataset (several hundred to several thousand records, with fewer than 20 fields in each record), you could expect NeuralSight to generate several hundred neural networks in around 1 hour or less, assuming a relatively current computer (vintage 2012, Intel i3 or equivalent with at least 4GB memory) and using NeuralSight default Build Options.

As a modeling strategy, a good approach when you begin a particular project is to configure a NeuralSight Session to:

  1. use whatever dataset you have;
  2. set the stopping criteria for 1 hour or 100 models;
  3. use NeuralWare suggested Build Options.

Let NeuralSight run to completion; that will give you an idea of how long it takes to train a representative mix of neural networks. Use the timing information to plan extended modeling Sessions.

Back to Top


How can I deploy a neural network created with NeuralSight?

NeuralSight is built on the Predict Engine, so it has the same deployment options as the Predict Engine. These include running the network in the Microsoft Excel environment with data in a spreadsheet, running the network using the Command Line Interface and data from a file, running the network in a custom application that is linked with the Predict Run-Time Kit and an appropriate data source, or compiling a FlashCode representation of the network and linking it with a custom VisualBasic, C, or Fortran 77 application. In addition, you can batch process new data files directly through the NeuralSight GUI. If NeuralSight is used to run prediction models, the average of the outputs from several models can be produced, as well as the output from each individual model.

If you want to deploy the network for use in an Excel spreadsheet that will be used on a computer other than the computer that created the network, a Run-Time Kit license and software are required for every computer that needs to run the network (in addition to having Excel on the computers).

If you want to deploy the network on another computer through an application that was created using a FlashCode representation of the network, a deployment license is required for every computer that needs to run the network. No additional software is required.

Back to Top


Professional II/PLUS Questions

What data limitations does Professional II/PLUS have?

Professional II/PLUS does not enforce any explicit data limitations on neural networks. The overall network size and the number of input fields allowed for a network are limited by available computer memory. Several hundred input fields for networks comprising 1 hidden layer and 50 hidden units are easily supported on vintage 2012 desktop or laptop computers.

How does Professional II/PLUS handle missing or invalid data?

Professional II/PLUS does not directly provide any data pre-processing facilities other than basic scaling (via MinMax Tables). You need to ensure that datasets used by Professional II/PLUS do not contain missing or invalid values. Note that NeuralWorks Predict data preparation capabilities can be used to prepare a dataset for use by Professional II/PLUS.

Back to Top


What types of neural networks can Professional II/PLUS produce?

Professional II/PLUS supports creating many of the named neural network paradigms referenced in neural network literature, including:

  • Feed-forward neural networks
  • Kohonen Self-Organizing Maps
  • Counter-propagation networks
  • Adaptive Resonant Theory (ART) and Fuzzy ART networks
  • Generalized Regression networks
  • Logicon Projection networks
  • Learning Vector Quantization networks
  • Modular networks
  • Probabilistic networks
  • Radial Basic Function networks
  • Reinforcement networks
  • Bi-directional Associative Memory networks
  • Boltzmann Machine networks
  • Hamming networks
  • Hopfield networks
  • Auto-associative Back Propagation networks
  • Recirculation networks
  • Adaline networks
  • Madaline networks
  • Perceptron networks

The Professional II/PLUS Instanet facility provides a guide for creating each type of network.

Back to Top


How many hidden layers and hidden units should I specify for a neural network?

Professional II/PLUS supports creating neural networks with more than 1 hidden layer; however, we recommend that you specify only 1 hidden layer. The Universal Approximation Theorem for neural networks provides theoretical justifcation for using only 1 hidden layer. From a practical perspective, a network with more than 1 hidden layer takes longer to train, and is more prone to over-fitting the training data.

Specifying the number of hidden units is always a challenge when creating a neural network manually. A reasonable starting point for the number of hidden units is 2/3 of the sum of the number of input variables and output variables: 2/3 * (inputs + outputs). After the network is trained, then you can try building a network with more hidden units, and another with fewer hidden units. Comparing the performance of all three networks will suggest whether to increase the number of hidden units or to decrease the number of hidden units.

Unless the training data has a very large number of inputs (more than 50), it is not advisable to have more than 50 hidden units (and usually you would specify fewer than 25 or 30 hidden units).

Back to Top


How long does it take to build a neural network with Professional II/PLUS?

The time required to build a neural network with Professional II/PLUS is very much dependent on the complexity of the network architecture you specify (the number of hidden layers and hidden units), the number of fields in a record and the number of records in the modeling dataset, and the number of training epochs (the number of passes through the dataset) you specify.

With current (2013) processor speeds in basic (mid-level) desktop or laptop computers running Windows 7 or 8, assuming 50 000 epochs for a dataset with several thousand records and no more than 20 fields in each record, one hidden layer, and fewer than 25 hidden units, Professional II/PLUS would typically produce a network within thirty minutes.

Back to Top


How can I deploy a neural network with Professional II/PLUS?

The easiest way to deploy a neural network built by Professional II/PLUS is simply to run it in Professional II/PLUS using new data from a file and writing network output to another file. With additional software development, you can create a UserIO program that permits more output flexibility.

You can also generate FlashCode for the network and compile and link the FlashCode into an application. If you want to deploy the network on another computer through the FlashCode application, a deployment license is required for every computer that needs to run the network. No additional software is required.

You can also use Designer Pack and create a custom application. This too requires a deployment license if you want to run the network on a computer other than the computer that has Professional II/PLUS and Designer Pack installed.