A Neural Network With Logical Reasoning Based on Auxiliary Inputs.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computer Vision

Background:

• Traditional neural networks often lack transparency in their decision-making processes.
• Explainable AI (XAI) aims to make AI predictions understandable through logical reasoning.
• Mimicking human deductive reasoning is a key challenge in developing trustworthy AI systems.

Purpose of the Study:

• To propose a novel neural network design incorporating auxiliary inputs for logical reasoning.
• To demonstrate how these auxiliary inputs can provide explainable predictions without sacrificing accuracy.
• To validate the effectiveness of this approach in real-world robotic applications.

Main Methods:

• Introduction of auxiliary inputs (indicators) to guide the neural network's reasoning process.
• Formulation of a reasoning process for cross-validation using these indicators.
• Testing the network design on MNIST dataset and robotic grasping image data.

Main Results:

• Auxiliary inputs, whether meaningful or not, can form the basis for explaining predicted outputs.
• Different sets of auxiliary inputs allow for diverse yet trustworthy explanations of the same outcome.
• Achieved 1-2% enhancement in prediction accuracy and removal of self-conflicting predictions.

Conclusions:

• The proposed neural network design with auxiliary inputs offers a flexible framework for explainable AI.
• This method enhances prediction reliability and provides interpretable reasoning, akin to human explanation.
• Potential applications include optimizing robotic learning in detection and grasping tasks.