Higher-Level Cognition - Computational Explorations in Cognitive Neuroscience - page 374

The Control of Activation-Based Processing The Nature of Activation-Based Processing

Representations

The first issue is one that we began to address in chap-
ter 9: How can the updating and maintenance of ac-
tive representations be controlled in a systematic, task-
relevant fashion? As explored in chapter 9, an active-
memory-based working memory system requires a dy-
namic control system. In this chapter, we continue to
explore the use of the dopamine-based gating system
developed in chapter 9. We see that this mechanism can
be used to drive a kind of trial-and-error search through
activation space, which can provide greater flexibility
than weight-based learning. As discussed in chapter 9,
it is likely that more complex models will require ad-
ditional control mechanisms, including a mechanism
based speculatively on the basal ganglia for dynami-
cally restricting the set of working memory represen-
tations subject to updating.
We emphasize the following important features of
these control mechanisms: (a) control is an emergent
phenomenon resulting from the interaction of basic
neural mechanisms (e.g., multiple constraint satisfac-
tion) with more specialized control mechanisms like the
dopaminergic gating system; (b) strategic, task-based
control develops through experience-based learning; (c)
control typically involves a hierarchical set of goallike
structures that decompose problems into a number of
simpler steps (e.g., the high-level goal is to make a sand-
wich, which constrains the lower-level goal of applying
peanut butter and jelly, which constrains the actions of
finding these ingredients) — goals can thus both con-
trol lower-level goals and be controlled by higher-level
goals.
In short, even though it can be useful to distinguish
control mechanisms from other parts of the system, this
distinction is often not a clear-cut one, and care should
be taken to treat the system as an interactive whole and
not as a master-slave kind of system; this later view in-
evitably leads to conceptions of controlling mechanisms
having homunculus-like powers. The reliance on learn-
ing mechanisms is also important for avoiding the need
to posit these homunculus-like powers — there is no
need for a homunculus if “intelligent” control can be
shown to emerge through experience.

The second issue for activation-based processing con-
cerns the development of the necessary representations.
How does a network like the activation-based catego-
rization one (figure 11.1b) get set up in the first place?
The general problem is that when the burden of process-
ing depends mainly on the strategic activation of appro-
priate representations (e.g., goals, discriminating fea-
tures, etc.), then there must be an existing vocabulary of
such representations available for immediate use. Fur-
thermore, one must be able to activate novel combina-
tions of such representations to solve novel tasks, which
is an essential aspect of many higher-level cognitive
phenomena.
To understand better the existing vocabulary con-
straint, consider the categorization task example — to
be able to switch rapidly among different categoriza-
tion rules by simply activating different representations
(e.g., “red,” “blue,” etc.), these representations have
to already exist and have appropriate associations with
corresponding posterior representations. It is not rea-
sonable to imagine that the necessary associations could
be developed dynamically on the fly — how could only
a few trials of learning develop appropriately rich and
systematic associations? Consider the analogy of trying
to have a reasonable conversation with someone in an
unfamiliar language — it is impossible to both learn a
large number of new vocabulary words and dynamically
use them for communicating. Even with a good, quick
dictionary, extensive experience is required to learn the
appropriate uses of words.
Therefore, the kind of slow, interleaved learning hy-
pothesized to be the specialty of the cortex (chapters 7,
9) may build up a rich vocabulary of frontal activation-
based processing representations, with appropriate as-
sociations to corresponding posterior-cortical represen-
tations. An activation-based control mechanism could
then rapidly select among an existing vocabulary. Nev-
ertheless, this general principle does not provide much
substantial insight into what kinds of things are actu-
ally represented in frontal cortex. Are they like “goals”
or “productions” or would simple copies of posterior
stimulus-based representations work just as well?