Page 286 - Deep Learning

Error Correction in Context 269

no way to count knowledge units (chunks, constraints, rules, schemas, etc.),
but rough estimates are available for some task domains. World-class chess
players might possess between 10,000 and 100,000 units of chess knowledge,
depending on their level of skill. The approximate size of the vocabulary of
13
the average, educated Western adult falls in the narrower range of 30,000 to
60,000 words. Large expert systems in the field of Artificial Intelligence con-
14
firm that human-level performance requires thousands of separate pieces of
knowledge. In Artificial Intelligence research, this is known as the Knowledge
Principle: Expert-level problem solving does not reside in the execution of
complicated reasoning algorithms but in the application of large amounts of
relevant knowledge. A third stream of information comes from work on the
15
implementation of intelligent tutoring systems, instructional software systems
that explicitly represent the subject matter that they teach. Systems of this sort
usually require several hundred knowledge elements to represent the com-
petence taught in an introductory course in, for example, algebra, computer
programming or database use. It is plausible that an expert knows 100 times
16
as much as a beginner, so the expert’s knowledge base is likely to be on the
order of tens of thousands of knowledge elements. The approximate conver-
gence of these four estimates provides modest confidence in the accuracy of
the 10k–100k range.
A computer program can acquire a knowledge base of this size via the
diligent typing of knowledge engineers, but human beings have to acquire
the knowledge through practice. Estimates of the learning rates of individual
learners are few and far between, but they range between 1 and 10 knowledge
units (chunks, rules) per hour of practice. If we assume a learning rate of 5
knowledge units per hour, and if we assume that the learner practices 4 hours
a day, 5 days a week, 50 weeks per year, then it takes 10 years to accumulate
50,000 knowledge units. In short, it takes years to become an expert because
the knowledge base of an expert is large in relation to the rate at which the
knowledge is acquired.

Domain specificity
Experienced physicians are better than residents at diagnosing medical
patients, but no better at troubleshooting electronic equipment than engineer-
ing students, and vice versa. A case could be made that the types of infer-
ences required in the two domains of medical diagnosis and troubleshooting
are similar at some level of abstraction, but few medical patients with chest
pains would feel comfortable being diagnosed by an electrical engineer. The
expert’s knowledge is encoded in memory in terms of the concepts that belong