Integrating Non-monotonic Logical Reasoning and Inductive Learning With Deep Learning for Explainable Visual Question...

Area of Science:

• Artificial Intelligence
• Cognitive Systems
• Machine Learning

Background:

• Deep network models dominate pattern recognition but require large datasets and computational resources.
• The "black box" nature of deep networks limits their application in critical, explainable domains.
• Existing methods struggle with explainability and efficient learning from limited data.

Purpose of the Study:

• To develop an explainable AI architecture integrating deep learning with commonsense reasoning and inductive learning.
• To address the limitations of current deep network models in terms of data requirements and interpretability.
• To improve performance on tasks requiring explanatory reasoning, such as Visual Question Answering.

Main Methods:

• Developed a hybrid architecture combining deep networks for feature extraction and answer generation with components for non-monotonic logical reasoning and decision tree induction.
• Integrated commonsense domain knowledge and incremental learning of unknown constraints.
• Evaluated the architecture on simulated and real-world image datasets and a simulated robot task.

Main Results:

• Achieved superior accuracy on classification tasks with small training datasets compared to end-to-end deep networks.
• Demonstrated comparable accuracy with larger datasets.
• Provided more accurate answers to explanatory questions and improved reliability and efficiency in simulated robot planning and diagnostics.

Conclusions:

• The proposed architecture offers a more accurate and explainable alternative to purely data-driven deep networks, especially in data-scarce scenarios.
• Integrating logical reasoning and incremental learning enhances the model's ability to handle complex, explanatory tasks.
• The approach shows promise for critical applications requiring transparent and reliable AI decision-making.