Neural Designer implements the most innovative artificial intelligence techniques.
Some of the main algorithms it contains are listed below.
Application types
- Approximation (or modeling) to discover intricate relationships.
- Classification (or pattern recognition) to recognize complex patterns.
- Forecasting (or time series prediction) to predict trends.
- Anomaly detection (using auto-associative networks) to flag unusual samples.
- Image classification (or image recognition) to identify visual categories.
- Text classification (or document classification) to categorize texts.
Data set
- Compatible with the most common data files: CSV, DAT, TXT, Excel and OpenOffice.
- Complete configuration of numerical, binary, categorical, date-time and constant variables.
- Complete configuration of training, validation, testing and unused samples.
- Random and sequential sample splitting.
- Missing values exclusion and imputation.
- Numerical and image data scaling.
- Descriptive statistics, distributions, box plots, time series plots and scatter charts.
- Variable importance using input and input-target correlations.
- Autocorrelations and cross-correlations for time series.
- Utilities for outlier detection and data filtering.
- Text data preprocessing.
- Image resizing and data augmentation.
Neural network
Network architectures
- Network architecture with unlimited layers.
- Dense networks for approximation and classification.
- Recurrent networks for forecasting.
- Convolutional networks for image classification.
- Auto-associative networks for anomaly detection.
- Attention networks for text classification.
Layers
- Dense layers for fully connected transformations and classification outputs.
- Convolutional layers for spatial feature extraction.
- Pooling layers for spatial and sequence summarization.
- Embedding layers for token representations.
- Multi-head attention layers for contextual representations.
- Long short-term memory layers for temporal dependencies.
Functions and transformations
- Dense activations: linear, hyperbolic tangent, rectified linear, sigmoid and softmax.
- Convolutions: same and valid, with configurable filters, kernel size and stride.
- Pooling: maximum and average.
- Embeddings with configurable vocabulary, sequence length and embedding dimension.
- Multi-head attention with configurable attention heads.
- Scaling, unscaling and clamping transformations.
Training strategy
- Loss functions:
- Mean squared error for approximation and forecasting.
- Mean absolute error for anomaly detection and robust regression; Minkowski error for robust regression.
- Normalized squared error for variance-normalized regression.
- Weighted squared error for imbalanced binary classification.
- Cross-entropy for binary and multiclass classification.
- 3D cross-entropy for language modeling.
- Optional L1 and L2 regularization for controlling complexity.
- Optimization algorithms:
- Adam and SGD for mini-batch CPU and GPU training.
- Quasi-Newton for full-batch tabular models on CPU.
- Levenberg-Marquardt for small sequential Dense models on CPU.
- Training control: automatic batch sizing, validation, stopping criteria, gradient clipping and best-model restoration.
Model selection
- Neurons selection algorithm for finding the optimal network architecture: incremental order.
- Inputs selection algorithms for selecting the most important features: growing inputs, pruning inputs, and genetic algorithm.
Testing analysis
- Testing errors, statistics, histograms and maximal errors.
- Goodness-of-fit analysis for approximation and forecasting.
- Outputs plots for forecasting.
- Confusion matrices for binary and multiple classification.
- Full metrics for binary and multiple classification.
- ROC curve and optimal threshold.
- Cumulative gain, lift and rates.
- Profit chart and misclassified samples.
Model deployment
- Output values and output data.
- Directional outputs for model exploration.
- Sample and model input importances.
- Response optimization under given conditions.
- Exportable mathematical expression.
- Exportable model in C, Python, JavaScript and PHP.
- Portable deployment packages for image and text classification.
Output
- Interactive reports with descriptions, tables and charts.
- Reports printable and exportable to ODT and PDF.
- Text, tables and chart data exportable to TXT and CSV.
- Tables, charts and network diagrams exportable to PNG.
Help
- Practical user’s guide.
- Extensive machine learning tutorials.
- Step-by-step solved examples in different fields.
- Frequently asked questions and video tutorials.
- Premium technical support by email, phone or video call.
Performance
- High-performance C++ implementation.
- Optimized memory management and numerical operations.
- Multicore CPU parallelization with OpenMP and Eigen.
- NVIDIA GPU acceleration with CUDA, cuBLAS and cuDNN.
Supported platforms
- Windows 64-bit.
- macOS for Apple silicon (M Series).
- Ubuntu Linux 64-bit.
Cloud computing
- Run Neural Designer on Amazon Web Services (AWS).