Page 227 - Deep Learning

210 Adaptation

procedure by weighing the reactants and the reaction products one has to
know that mass is conserved in chemical reactions.
In general, prior declarative knowledge about the task environment is
the learner’s main resource for judging the appropriateness of the situations
he or she creates or encounters in the course of performing a task. Learners
monitor their own actions by comparing what they observe with expectations
based on what they know. Errors appear subjectively as conflicts between what
the learner knows ought to be true and what he or she perceives to be true. Prior
knowledge implies that such-and-such should be the case, but the actual situ-
ation fails to conform; hence, some action taken en route to that situation was
incorrect.

Dissociation Between Action and Judgment
The hypothesis that learners detect errors through prior knowledge is para-
doxical: If the learner has enough knowledge to recognize a particular action
as incorrect, then why did he or she perform that action in the first place? Why
not apply that knowledge so as to select the correct action? On the other hand,
if the learner does not have sufficient knowledge to produce the correct action,
how can he judge his action as incorrect? The ability to catch ourselves making
errors points to a duality in our relation to the environment.
The dissociation between execution (action) and evaluation (judgment)
can be understood in terms of the distinction between declarative and practical
knowledge. This Dual Knowledge-Base Principle is a well-established aspect of
11
human cognition. The function of practical knowledge is to generate action in
the service of goals. The function of declarative knowledge is less obvious. The
functions most frequently proposed are to enable people to describe and explain
phenomena, infer new conclusions and predict future events. Declarative
knowledge is said to consist of propositions. A proposition is the meaning of a
statement or an assertion. It is meaningful to ask whether a proposition is true
or false, even though establishing the answer is sometimes difficult. This view of
declarative knowledge was invented and developed by philosophers, linguists
and artificial intelligence researchers for their various purposes.
However, this view leads to conceptual riddles that have befuddled think-
ers in all those disciplines. After two millennia, there is no widely accepted
12
theory of what it means for an assertion to be true. We can understand the
truth of a specific proposition like the cat is on the mat in terms of a correspon-
dence between assertion and reality, but this interpretation of truth is already
problematic with respect to such innocent-looking assertions as all swans are