Page 264 - Deep Learning

Error Correction: The Specialization Theory 247

53
I propose that the relation between instruction and learning is similar.
According to the Information Specificity Principle introduced in Chapter 6, the
learning mechanisms in a person’s brain take particular types of information
as input. Assisted learning is more efficient than unassisted, not because there
are dedicated brain mechanisms for assisted learning, but because the instruc-
tor arranges for information of the relevant types to be more abundant than
they otherwise would be, thereby making the learning mechanisms in the stu-
dents’ brains do more work, or different work, in a given period of time.

Simulating Learning From Tutoring
If this line of reasoning is correct, a model of unassisted learning is also a
model of assisted learning and it should be possible to turn the underlying
learning theory into a set of design principles for effective instruction. But in
the computer simulations of constraint-based error correction discussed pre-
viously in this chapter, the relevant constraints were assumed to be acquired
before practice began. In the counting simulation, the constraints are hypoth-
esized to be part of the child’s innate cognitive structures. In the case of col-
lege chemistry, the constraints must be acquired while the student studies the
relevant textbook chapters. In both simulations, the entire set of constraints
was available to the model at the outset of practice. In those scenarios, the HS
computer model acted as a model of unassisted skill acquisition.
According to the Piggyback principle, the HS model should not need to
be extended with any additional learning mechanisms to model assisted learn-
ing. When the learner does not know all the relevant constraints, he cannot
recognize all of his constraint violations. Learning opportunities are missed.
But a tutor can recognize the constraint violations that the student overlooks
and alert the learner to them. By calling attention to an error and supplying
the learner with the information that is needed to correct it, a tutor makes the
constraint-based specialization mechanism do more, and different, work than
it would have done otherwise. The interaction with the tutor allows the learner
to operate as if he knows more than he does, because his own constraint base
is augmented, functionally speaking, with the more extensive constraint base
of the tutor. All that is needed to model learning from tutoring with HS is to
reinterpret the constraints as feedback messages provided by a tutor in the
course of problem solving instead of knowledge structures retrieved from the
learner’s long-term memory.
To verify that the HS simulation model passes this test Andreas Ernst,
Rees and I tutored HS in subtraction with regrouping, an instructional topic