Curriculum for the Machine
Copyright (c) 2002 Jeff D. Chapman
Electronic Rights granted to FleaByte; All Other Rights Reserved
Much discussion has centered on the feasibility and implementation of Natural Thinking Machines (NTMs, also known as "intelligent machines") -- computers that have the capabilities to think like humans [1-3]. Various paradigms have been presented for enabling people to design such machines . In this essay I use the phrase Natural Thinking Machine to characterize a computer with sufficient hardware and software to "learn." More than just storing information, in this sense for a machine to "learn" it also needs to demonstrate:
Little has been discussed about what we should expect from NTMs or how we should go about training such machines. It is fine, for example, to say that if we are designing a computer program to play chess we should run this certain set of historical examples to assist the machine in learning successful strategies. But what if our goal is to design a computer program with many of the capabilities of a human? What should it be taught? How do we go about setting metrics to measure success? What would be the appropriate curriculum? This paper discusses some approaches for designing Curriculum for the Machine and raises some of the philosophical issues inherent in any machine curriculum.
Do we Need a Curriculum?
When a machine demonstrates sufficient skills to "learn" then we should begin to approach the process of "teaching" it in the same fashion that we approach the process of teaching children. Hence we should begin discussing a curriculum.
A curriculum is a set of instructional materials and methodology for teaching; it is a framework that bolsters the transfer of knowledge. More than setting expectations for what should be learned, a curriculum is also an artifact of a culture that conveys many of the values of that culture to the recipient . Often times the curriculum teaches more than what we think the recipient should know: it also teaches how we want them to behave.
A curriculum does not need to be formally defined; for
example most parents teach their toddlers a tacit curriculum regarding
personal safety: don't touch the stove, don't eat the dirt, and don't
poke things into the electric socket. A formal curriculum is typically
specified when we want to present our teaching materials or methodology
to administrators, teachers, or potential students for one purpose or
another. Some reasons to present a curriculum include:
Given the inevitable future existence of Natural Thinking Machines, curricula might potentially be useful for the same reasons that we develop curricula for humans. For example, at some future time organizations may emerge that certify NTMs -- it might be useful to know if your NTM has earned its PhD in physics. As society may have different expectations for NTMs than humans, the curriculum required to grant an NTM a PhD might be different than that required for a human.
There might additionally be reasons to prepare a curriculum for a machine beyond those described for humans. For example, if an NTM needs to be repaired or have a software upgrade installed then it might be desirable to review its functions to assure that it still conforms to the knowledge base and behavior of another NTM given the current curriculum.
Faster hardware and better software will promote more capable machines that may eventually “learn” and will be able to be taught. Since putting our head in the sand and saying "this shouldn’t be done" doesn’t dissuade people from actually doing it, we had best start thinking about HOW we should teach machines, or how we should regulate what they are taught.
Components of Curriculum
The components of a curriculum address the wide variety of our learning styles. A curriculum for instruction toward a Bachelors Degree in Mathematics might include listening to lectures, reading textbooks, research projects, software models, writing assignments, group discussions, and problem solving. We devise components to match how a human being learns: by reading, by listening, by writing, by thinking through problems, and by sharing ideas with others .
A Natural Thinking Machine may however learn differently than a human; the components of its curriculum might for example stress interaction with other NTMs over the Internet, black box testing, populating knowledge databases, and simulations. It is rather probable that teachers may transfer knowledge to an NTM differently than how they transfer knowledge to another human. The curriculum will therefore need to be modified to accommodate the NTMs learning modus operandi.
Who Determines the Curriculum?
The objectives, patience, and capabilities of teachers determine the curriculum along with the needs of society. Battles often rage however over the nature of federal versus local control of the curriculum. Even in organizations involving only humans this is a political "hot potato." Throw machines into the mix and you open the floodgates to speculations and new concerns.
The creators of the Natural Thinking Machines (the corporate "parents") may certainly have a strong say in the matter, choosing to allow their machines to learn only from approved sources. After an NTM is sold the new owner may have more local control over the curriculum. This is a contentious issue, even with the most rudimentary NTMs such as Aibo . The owner may choose how and when to train an NTM and might be interested in reviewing the effectiveness of various curricula. At some point an NTM might become "self-actualized" and choose its own curriculum.
As NTMs would probably have a major impact on society we can expect the battle for oversight of their curriculum to rise to the highest levels. We can expect both accreditation and government agencies to introduce regulation of NTM curricula.
How Society Enforces the Curriculum
We elect or appoint Boards of Education to review and approve human curriculum. It is probable that society will create similar administrative bodies to review and approve Natural Thinking Machine curriculum. People serving on an NTM Board of Education, however, will have somewhat different talents than those that sit on their human counterpart. An NTM Board of Education member will need to understand the technology, the rate that technology changes, and the differences between various popular NTMs.
Teaching institutions usually seek accreditation for their curriculum. An accreditation agency evaluates the curriculum and the student diplomates to verify that they have met the curriculum's objectives. An accreditation institution for NTM teachers may, however, operate more like a cross between Good Housekeeping and the Underwriters Laboratories -- vouching both for the quality and the safety of the NTM after it has received the curriculum. Accreditation criteria frequently change ; as new generations of NTMs progress, a teaching institution may need to constantly update its curriculum to maintain accreditation.
Professional organizations abound to certify that human beings have met the education standards described by a curriculum. Many base their certification upon the completion of courses at an accredited institution, as well as allowing the recipient to qualify by standardized testing. We can expect similar outfits to gain respectability for certifying, for example, that your NTM butler meets its high standards. The value of a NTM may be based upon what certifications it holds as well as its intrinsic learning capabilities. Laws will be passed regarding misrepresentation of certification; laws may forbid the sale of some NTMs for critical purposes unless they have been properly certified, similar to the process that the FDA now uses for medical devices .
What should be our objectives for determining the curriculum for a Natural Thinking Machine and how might this differ from our objectives for a human curriculum? Our NTM curriculum should augment the NTM to assist us and generally improve our lives, not create unemployment and make us miserable! NTMs may be built with certain skills for various specific objectives -- the context of how they might be used plays a role in determining curriculum ; this becomes focalized by specialization.
We might have opportunities to utilize NTMs to satisfy
human requirements in the areas of survival, procreation, pleasure, vocation,
education, discovery, work, and science. For example, we could imagine:
When training a NTM to cook meals the context is perhaps the major determinant of our curriculum! (What does the owner buy at the supermarket and how does he make that decision? What are his eating habits? What facilities does he have available for preparing meals? How long do various foodstuffs last?) The curriculum is largely established by the context of the owner’s neighborhood, apartment, and tastes. In a sense the context delimits the curriculum -- we don’t have to include as much in the curriculum if we know the context ahead of time. Hence, an NTM might earn its official approval of "certified chef" for Los Angeles, but be unauthorized to cook in London. As an NTM progresses through its education it might gradually narrow its curriculum by increasing its specialization much as we humans do (learning to read English is rather broadly focused in second grade; in-silico simulation of protein folding is rather narrowly focused when we earn our PhD).
Should the curriculum include softer philosophical objectives ("balance" perhaps; maybe "global outlook")? After all, if we are training a machine to learn how to promote human objectives, we need to be wary to avoid "ends justifying the means" -- the machine needs to visualize the impact of its actions upon the long-term well being of the world.
Furthermore a machine capable of learning also needs to be taught some of the most basic human considerations that most of us learn at a very young age, such as courtesy and to avoid hurting others. Any curriculum may need to stress positive non-violent behaviors even to the extent of going out of the way to avoid any accidental damage or injury to humans or their possessions. Much of what we take for common sense in humans, or the evolved sense of family, might take a machine a couple of years to learn.
Managing Changes to the Curriculum
Quite unlike humans each successive generation of Natural Thinking Machines might potentially increase their learning potential geometrically. Part of the challenge of "teaching" these machines may then be to tailor the curriculum for the given model.
Machine-learning technology advances both by hardware improvements and in paradigm shifts affecting the design of the software. The curriculum we use to teach a Compaq 2030 Entrepreneurial model G with MicroLearn release 7.3 installed might be different then what we use for other computers, other learning software programs, or even other versions of MicroLearn. Part of teaching any curriculum, however, is first assessing the capabilities of the student. For NTMs then, teachers need to assess what curriculum works best for the available array of hardware and software. The contents of the curriculum however may be quite similar even though the methodology of "teaching" might vary. Following the above example, we want to teach the machine how to avoid burning the toast, regardless of the machine model.
Natural Thinking Machines have the potential to transform our relationship with computers into an ever-changing arena of human teachers with silicon students. Societal pressures will demand curriculum for teaching NTMs to certify their capabilities. NTM curriculum is likely to vary from that designed for human counterparts and should consider the context of NTM use, ongoing changes in the technology, and safeguards to humans.