NeurEco Tabular#

For Tabular templates both input and output features are sets of scalar parameters.

Tabular is proposed with:

  • Graphical user interface

  • Python API

  • Command line interface

NeurEco Tabular requires a neureco_tabular license. This license gives access to all functionalities for Tabular except for the export of the model to the formats other than the native binary files.

Many NeurEco features were developed with a view to increase even more the model’s embeddability:

For user’s flexibility all these features are made available under neureco_tabular license. An add-on license is required only when preparing for actual embedding of the NeurEco model.

An add-on license neureco_tabular_embed is a license specific for export to FMU, ONNX, C head file and VBA formats (see, for example, Export NeurEco Regression model with the GUI).

There are currently three tabular templates.

Tabular Regression#

This template addresses most of engineer’s needs. It is used to create neural network models where the outputs are continuous variables, corresponding to the modeling of physical phenomena. NeurEco will create a regression predictive model that approximate the underlying process by a function \(Y = f(X)\), where \(X\) is the input and \(Y\) is the output of the model.

For a quickstart tutorial, see NeurEco Tabular Regression quickstart tutorial

For a full documentation, see Tabular Regression

Tabular Classification#

This template is used to create neural network models for classification problems where the output describes the belonging of the input to one of the few possible classes.

Tabular Regression meets most engineering needs. For example, a binary response indicating that a system is defective occurs when its response exceeds a given threshold of stress, temperature or/and number of cycles… To solve most classification problems, it is better to post-process the output of a regression model by applying a threshold. This approach can result in better classification because the targeted physical value is rich with information contrary to the classification poor binary target.

We propose this Tabular Classification template because it is not always possible to reformulate the problem into a regression problem or to get the corresponding physical data.

For Tabular Classification the model is still \(Y = f(X)\), but \(Y\) is a binary one-hot encoded vector of at least two components: one of the components of \(Y\) is set to one, while all the other components are set to zero. If the \(i^{th}\) component is set to one, it means that the corresponding \(X\) belong to class \(i\).

For a quickstart tutorial, see NeurEco Tabular Classification quickstart tutorial

For a full documentation, see Tabular Classification

Tabular Compression#

This template is a powerful tool to reduce dimensionality of \(X\), when \(X\) is a large vector. This reduces the size of data and it complexity.

In this case we create two neural networks, a compressor \(C\) and a decompressor \(D\). The input vector \(X\) is compressed to \(x\), such that \(x=C(X)\), where the dimension of \(x\) is small in comparison with the dimension of \(X\). Moreover, it is possible to reconstruct \(X\) using \(X=D(x)\).

The main application of this compression process is to use \(x\) instead of \(X\) as the input to a regression model.

For a quickstart tutorial, see NeurEco Tabular Compression quickstart tutorial

For a full documentation, see Tabular Compression