I wanted to share some thought-provoking ideas from a recent talk that I attended on "Building AI with Common Sense", presented by Ernest Davis, professor of Computer Science at New York University. His research focuses on the automation of common sense, spatial, and physical reasoning.
The image below really sets the tone for the topic. Can you see what's happening?
Amazon Image Recognition processed this image and is 98% confident that it sees a "Robot" and a "Toy". Those are correct, from a limited sense. However, do you see what the robot is doing? The common sense aspect of this image is probably the most important part of it!
Most of the AI that we're familiar with centers upon machine learning, statistical based learning, sourced from training sets. This has demonstrated significant advances in recent years, especially from neural networks and deep learning. However, the scope of these advances remain very narrow, focused on specific domains and tasks (analyzing sentiment, classifying movies, identifying cats vs dogs, speech recognition). Even with recent advances, AI is extremely poor at generalizing across concepts.
The talk began by stating, "commonsense knowledge and the automation of commonsense reasoning is one of the greatest challenges in artificial intelligence".
So, what's already been tried?
- Hard-Coding Expert Knowledge
- Cyc Project (the longest running AI project)
- Specialized Ontologies and Taxonomies
- Web Mining
- Never-Ending Language Learning (NELL project)
- Crowd Sourcing (ConceptNet, Amazon Mechanical Turk)
- Machine Learning
The above techniques are helping to grow our ability to use common sense knowledge. However, they fail to generalize across domains.
The presentation concluded with some ideas on how to move forward. Techniques such as creating new foundational theories of common sense reasoning, applying common sense to the meaning of text, utilizing structured symbolic representation of logic, and developing innate knowledge.
These are definitely some intriguing ideas to consider, especially the idea of applying common sense AI to the natural language processing (NLP) that we do on a daily basis for sentiment, news, entity extraction, and more.
The presenter has a book coming out later this year, titled, "Rebooting AI: Building Artificial Intelligence We Can Trust". Feel free to check it out. Thank you.
Kory Becker
