PV3 ⊧ NNRepairrepairs neural network classifiers based on examples

Constraint-based technique to repair neural network classifiers

Application domain/field

• Neural networks
• Neural network classifiers
• Program repair
• Fault localization
• Faulty parameters
• Constraint solving

Type of tool

Program (neural network) repair

Expected input

• Neural network used for classification
• Positive and negative examples, i.e. dataset for classification

Format:

• Neural network: Keras model
• Dataset: ?

Expected output

Repaired neural network.

Internals

NNRepair targets neural networks used for classification. First NNrepair uses fault localization to find potentially faulty parameters (weights on edges connecting neurons) in a neural network. Then it performs repair using constraint solving to apply small modifications to the parameters to remedy the defects. The tool is based on NeuroSPF. It uses Symbolic PathFinder and Z3.

Comments

A combination of finding a problem (PV1) and suggesting how it can be solved (PV2)