Sensitivity analysis for Tabular solutions#

The Sensitivity analysis for Tabular solution provides the sensitivity of any output feature to the state of any node in the network, including the nodes of prime interest: input features.
This functionality is based on the input perturbation algorithm, the general idea of which is:
  • Add a set of perturbations of different magnitudes to the node under consideration

  • Calculate the corresponding variation of the output

  • Repeat the process for each node independently

  • Deduce the common rank of importance of nodes

The Sensitivity analysis give an insider look at the model’s logic and can sometimes reveal that some input features are more important than others. This can lead to changes in the choice of inputs features. If one of input features has a little impact on the outputs, it can be excluded from training. The new training with less input features generally leads to a smaller model without significant loss of accuracy.

The NeurEco Sensitivity analysis is performed in Network sensitivity sections of GUI and proposes two modes:

Sensitivity analysis for a single sample#

  • Switch to Evaluation panel

  • Choose the file in Evaluation files section:

    • If the file was supplied earlier, it is already listed in Evaluation files

    • To add new file for evaluation, press + in Additional section of Evaluation files

    • For Sensitivity analysis, the output file is not required

  • Once the input file clicked (or a pair input/output), choose a sample to study using Sample slider

  • Click on one of the output neurons (representing output features) on the plot of the neural network in the Network sensitivity section

  • Each neuron becomes colored according to the sensitivity of the chosen output neurons with respect to this neuron

  • For the input neurons (representing the input features): click on an input neuron to get the calculated value of sensitivity in addition to color

An example of the sensitivity analysis for a single sample:

Tabular network sensitivity for a single sample

Tabular network sensitivity for a single sample. Regression test case: Energy consumption.#

Note

By default, the Network sensitivity is performed with the last model available in the checkpoint. Use the checkpoint slider in the bottom to choose any other available model.

Sensitivity analysis for a whole dataset#

The sensitivity for a whole dataset is calculated as a norm of sensitivities on each sample in this dataset.

  • Switch to Metrics panel

  • Choose the file in Evaluation files section:

    • If the file was supplied earlier, it is already listed in Evaluation files

    • To add new file for evaluation, press + in Additional section of Evaluation files

    • For Sensitivity analysis, the output file is not required

  • Click on one of the output neurons (representing output features) on the plot of the neural network in the Network sensitivity section

  • Each neuron becomes colored according to the sensitivity of the chosen output neurons with respect to this neuron

An example of the sensitivity analysis for a single sample:

Tabular network sensitivity for a whole dataset

Tabular network sensitivity for a whole dataset. Regression test case: Energy consumption.#

Note

By default, the Network sensitivity is performed with the last model available in the checkpoint. Use the checkpoint slider in the bottom to choose any other available model.