Human-Assisted Training for Artificial Intelligence

Introduction to HAT-AI

By James K. Baker^1,2,3, Bradley J. Baker^1,4, Xuedong Huang⁵, Raj Reddy³, Tom Mitchell³, Ivan Garibay², Bhiksha Raj³, Rita Singh³ and Michael Georgiopoulos²

Artificial intelligence, particularly deep learning with neural networks, has had many dramatic successes in recent years. However, these dramatic demonstrations obfuscate the fact that there are still many large gaps between the capabilities of current AI systems and true intelligence. In addition, the research benchmarks by which the relative performance of AI are measured also fail to measure these important gaps. Indeed, some common criteria for determining whether a machine learning system exhibits "intelligence" also ignore these gaps. For example, one criterion that has been used for many decades is that if machine learning system can do as well as a human on a task that indicates intelligence when done by a human, then the machine learning system is said to exhibit "intelligence." This criterion has been met by chess-playing programs for several decades. In the current decade, it has been met on a wide variety of tasks, including many real-world applications.

Revised Criteria for Machine Intelligence

The gaps in the above criterion of "machine intelligence" are shown by the following more ambitious criteria:

• Sensibility: An intelligent system should make "sensible" decisions:
  1.      In pattern recognition, a mistake is not sensible if a human would say "That's stupid! No one would make that mistake."
  2.      Such mistakes are common, for example, under adversarial attacks.
  3.      A decision boundary is not sensible unless a human would agree that the classification of a pattern on the boundary is uncertain.
  4.      Evaluating sensiblity requires human judgment. We are deliberately specifying a criterion that requires a human to evaluate it.
• Socratic wisdom: An intelligent system should "know what it knows and what it does not yet know."
  1.      The system should be able to estimate the likelihood that it is wrong.
  2.      The system should also be able to evaluate the reliability of internal variables, such as activations of inner nodes in a DNN.
  3.      Socratic wisdom is related to the ability for introspection, another indicator of intelligence.
  4.      Introspection and Socratic wisdom may be implemented use judgment nodes.
• Interpretability: At least some of the decisions and the inner workings of an intelligent system should be interpretable by a human observer.
  1.      This criterion also requires human evaluation.
  2.      More generally, an intelligent system should be able to communication its knowledge to other intelligent systems.
  3.      As with humans, a large part of the inner workings may be at an "unconscious" level, not available to be communicated.
  4.      Intrepretability may be enhanced by using knowledge sharing.
  5.      Interpretabilty may be enhanced by using judgment nodes.

The Future of Machine Learning

To make these goals possible, we are proposing research that breaks the number one unwritten rule of AI research: "hands off during training." Indeed, it will be very difficult to meet the criteria listed above without human assistance. The concepts of "interpretability" and "sensibility" are defined in terms of human reaction. Even humans often fail to have Socratic wisdom, as Socrates himself pointed out. However, intelligent people meet all these criteria to some degree and current machine learning systems generally do not.

More specifically, we propose systems in which humans and AI systems cooperate in the training of the human+AI systems. We call this methodology "Human-Assisted Training of Artificial Intelligence" or HAT-AI for short.

From another perspective, HAT-AI represents a radical change in the long-term direction of AI. Rather than envisioning a future in which autonomous AI systems first achieve Artificial General Intelligence and then eventually achieve Artificial Super Intelligence, humans and machines should begin working cooperatively immediately.

A cooperative human+AI system changes the goal for artificial general intelligence. As a minimum, a cooperative human+AI system should at least match the performance of either a human working alone or an artificial system working alone. In other words, a fully functioning cooperative human+AI system should have general intelligence since a human working alone has general intelligence. This point of view of cooperative intelligent systems gives a very different, less dystopian, vision of future systems with general and then super intelligence.

Human-assisted training for artificial intelligence is just one, although perhaps the most controversial, of many proposals within the broader concept of "cooperative AI." Machine learning as part of a cooperative team of humans and computers is one of the methods of machine learning described by Mitchell as The Future of Machine Learning. This website hopes to be a bridge between the present and that future. Other concepts in cooperative AI include ethical and moral considerations and many other issues as well many other tools and approaches in addition to human-assisted training.

Additional Desirable Properties for the Future of Machine Learning

1. Lifelong learning, also called continual learning
2. One shot learning and other methods of learning with little data
3. Self-supervised learning
4. Insightful generalization -- inferring a general pattern from a few examples
5. Adaptability, an application of continual learning to changing conditions
6. The ability to communication information and knowledge to other intelligent entities, both human and machine
7. More to come

Intelligent humans have these capabilities, at least to some degree. A goal of cooperative human + AI systems is to do even better. HAT-AI can help.

Desirable Properties of the Training Process

1. AI-assisted automated development testing
2. A human + AI learning management system that is itself trained by human-assisted training
3. More to come

Broader Discussions of the Future of Machine Intelligence

This website specializes in human-assisted training for AI, with special emphasis on techniques that help meet the proposed revised criteria for machine learning. Many of the topics presented here are specific to human-assisted training of deep neural networks. However, some discussions are much broader than that. Some of the ideas apply to differential programming in general, not just to neural networks. Some of the ideas apply to neural networks but break the rules for automatic differentiation, so they extend neural networks in a different direction than differential programming. Some of the ideas apply to many other types of machine learning algorithms.

Human-Assisted Training is Beneficial but Not Essential

Some of techniques proposed here use a cooperative human + AI learning management system. In some implementations, a fully automatic AI-controlled learning management system could be used instead. These implementations would then not be instances of "human-assisted training" for AI, so the human assistance may be beneficial but not essential.

Cooperative systems with teams of human and machines are desirable even if the humans do not participate in the training process beyond the initial design. This website distinguishes between "human-assisted training for AI" and "human-assisted AI", in which humans do not necessarily participate in the training. Research in human-assisted AI is very worthwhile and several groups are pursuing it. Although human-assisted AI is not the focus of this website, some of the techniques presented here contribute to human-assisted AI in general, not just of human-assisted training.

HAT-AI is a New Area with Many New Ideas to Explore

Presenting New Ideas in a Deliberately Informal Style

This website is a work in progress. It is expected to grow like a many-branched blog. Contributions are welcome. Anyone who wishes, whether student, faculty, researcher, amateur experimentor or interested observer may become an Evangelist for HAT-AI.

Many of the ideas and suggestions on this website will be presented here for the first time, with a style more akin to a blog or to brainstorming than to finished resarch publications. This style is due in part to the current status of human-assisted training for AI. It is a new field that breaks rules and changes the criteria in the definition of machine intelligence. The informal style and release before formal publication is also a deliberate attempt to enable students to do work at the cutting edge. To further facilitate work by students, many of the projects to be suggested here will be designed to be done with limited computing resources.

Project Ideas for Students

There will be an attempt to give students priority in working with new ideas in human-assisted training for artificial intelligence. Registered students may be provided early access to ideas and suggestions in topics in which they have expressed an interest. A student's faculty adviser will be included in the communication at the student's request. A student will be put in contact with other students working on the same or similar topics if the students mutually so request. At a student's request, the student will be put in contact with a mentor, if one is available. A mentor may be a faculty member at the student's school or at another institution. A mentor also may be another student, a researcher at a government or non-profit research lab, a researcher working for a commercial company or a retired researcher.

Researchers at large research labs are asked to help us give priority to students. Please consider volunteering to be a student mentor. As a mentor, you will share early access to selected material. You may also mentor projects based on topics that have been posted on this website or other topics in human-assisted training for artificial intelligence. Please help the students you mentor to be successful contributors and co-authors. Try to guide the students to be the primary contributors.

To Register

Students: To register, send a message to "students at-sign hat-ai.com". Tell us about your background and topics of interest. Let us know if you are already working with a faculty member or other mentor. If you request, we will try to put you in touch with a mentor. Please let us know if you would like to be put in contact with other students.

Faculty: To register, send a message to "faculty at-sign hat-ai.com". Tell us your background and interests. Introduce us to any students working with you who are interested in HAT-AI. Please let us know if you are willing to mentor students from other institutions.

Mentors: If you wish to volunteer to be a mentor, send a message to "mentors at-sign hat-ai.com". Tell us about your areas of expertise and the topics for which you are volunteering to mentor.

Contributors: Contributions are welcome. You may contribute informal blog-like entries, longer essays, or finished research papers. Include your name, co-authors, and affiliations. You may include your own copyright notice if you wish. Send contributions to "contributors at-sign hat-ai.com".

Evangelists: Spread the word about Human-Assisted Training for Artificial Intelligence! Recruit faculty, students, mentors and other evangelists. Introduce yourself to us at "evangelists at-sign hat-ai.com" and send us frequent updates on your activities. Thanks!

Don't Worry about Empty Links

Don't worry about empty links on this website. They may have been left empty on purpose. An empty link may be due to a future topic that has been withheld from posting to provide early access to students.

Don't worry about "Coming Soon" messages either. Such a message may be used when there is a reference from another page to a topic that is planned but not yet finished, or that is not yet ready for release. The release of a topic may be delayed because it is withheld to allow early access to students or because it is the subject of an active brainstorming effort. These "Coming Soon" messages are a symbol of the active on-going development of ideas for Human-Assisted Training for Artificial Intelligence.

Affiliations:

1. D5AI LLC
2. University of Central Florida
3. Carnegie Mellon University
4. University of Massachusetts
5. Microsoft

Navigation:

• Home
• Human + AI Learning Management System
• Sensibility
• Introduction to Knowledge Sharing and Interpretability
• Judgment and Socratic Wisdom
• Experimenters' Page
• More to come

© D5AI LLC, 2020

The text in this work is licensed under a Creative Commons Attribution 4.0 International License.

Some of the ideas presented here are covered by issued or pending patents. No license to such patents is created or implied by publication or reference to herein.