Key features

• Create and apply neural networks to:
  • Forecasting
  • Classification
  • Function Approximation
  • Data Anomalies Detection
• Analyze and preprocess datasets
• Automatically search for the best neural network architecture
• Analyze network performance with graphs and detailed statistics
• Easy-to-use interface

Analyze and Pre-process Your Data

• Import Excel files
• Import popular ASCII file formats (CSV, TXT, PRN)
• Custom date formats and file structure definition
• Input dataset size is limited only by the hardware of the computer


• Date/Time values encoding
• Categorical values encoding
• Numeric values scaling
• Min/max values specification for numeric columns scaling


• Missing values handling for both numeric and categorical data
• Outliers handling for numeric data
• Automatic recognition of data entry errors (wrong type values)
• Visual representation of data anomalies in the Dataset window


• Automatic and manual column type identification (numeric, categorical, date, time, text)
• Random, sequential and manual dataset partition onto training, validation and test sets
• Accept/ignore records and columns manually


• Statistical information for data columns
• Binary columns for anomalies indication
• Two methods of automatic lag columns insertion
• Preprocessed data representation
• Detailed Data Analysis and Data Preprocessing Reports

Design Neural Network

• Input feature selection (GA, stepwise, exhaustive).
• Manual architecture specification (up to 5 hidden layers for multi-layer perceptron)
• Heuristic architecture search with customizable range of search and sensitivity
• Exhaustive architecture search
• Customizable search range and search sensitivity
• Detailed statistics for each tested architecture
• Network fitness criteria: AIC, Test set error, Correlation, R-squared
• Graphical representation of network fitness
• Time-series networks
• Network sets
• Network visualization


• Training algorithms: Conjugate Gradient Descent, Levenberg-Marquardt, Quick-Propagation, Quasi-Newton, Quasi-Newton (Limited Memory), Incremental and Batch Back-Propagation
• Automatic adjustment of learning rate and momentum for Back-Propagation algorithm
• Activation functions: Linear, Logistic, Tanh, Softmax
• Error functions: Sum-of-Squares, Cross-entropy
• Classification model: Winner-takes-all, Confidence-limits (Accept/Reject levels)

Control Network Training Process

• Real-time training error graph
• Real-time control on training parameters:
• errors on training and validation set: MSE, MAE, CCR
• error improvement
• training speed (iterations per second)
• # of iterations.
• Continue training with new parameters
• Jog weights
• Add jitter


• Correlation and r-squared real-time graphs
• Error improvement graph
• Weights distribution graph
• Error distribution graph
• Input importance graph
• Training log: test and validation set error for each iteration


• Early-stopping on generalization loss
• Retain and restore best network
• Stopping conditions:
• target error on training and validation sets: MSE, MAE, CCR
• error improvement: network error, dataset error
• number of iterations
• generalization loss
• Automatic network retrains and selection of the best network among retrains
• Retrains statistics
• Weights initialization: manual randomization range; optimized for Uniform or Gaussian distribution

Test and Analyze Performance

• Actual vs Output graph
• Scatter plot
• Response graph
• Confusion matrix
• ROC curve
• Actual vs Output Table with absolute and relative errors
• Input importance graph

Apply Network

• Enter new cases manually or insert from the Clipboard
• Load new cases from a new data file
• Apply to selected records from your original dataset
• Graphical network output representation
• Output representation with Results Table
• Confidence limits for network output
• Save results in a separate file or copy them to the Clipboard

General

• Customizable interface
• Detailed reporting
• Online help system
• Free technical support
• Project files to keep all related information in one place
• Sample financial, marketing, real estate and scientific problems included