Combined Model and Task Learning, and Other Mechanisms - Computational Explorations in Cognitive Neuroscience - page 176

H0 H1 H2 H3 H4

act in a positive way to produce better overall learn-
ing. In the next section, we will see that although bidi-
rectional connectivity can sometimes cause problems,
for example with generalization, these problems can be
remedied by using the other principles (especially Heb-
bian learning and inhibitory competition). The subse-
quent section explores the interactions of error-driven
and Hebbian learning in deep networks. The explo-
rations of cognitive phenomena in the second part of
the topic build upon these basic foundations, and pro-
vide a richer understanding of the various implications
of these core principles.

V0 V1 V2 V3 V4

Output

Hidden

Input

Figure 6.3: Network for the model-and-task exploration.
The output units are trained to represent the existence of a line
of a particular orientation in a given location, as indicated by
the labels (e.g., H0 means horizontal in position 0 (bottom)).

6.3 Generalization in Bidirectional Networks

One benefit of combining task and model learning in
bidirectionally connected networks comes in general-
ization . Generalization is the ability to treat novel items
systematically based on prior learning involving similar
items. For example, people can pronounce the nonword
“nust,” which they've presumably never heard before,
by analogy with familiar words like “must,” “nun,” and
so on. This ability is central to an important function
of the cortex — encoding the structure of the world so
that an organism can act appropriately in novel situa-
tions. Generalization has played an important role in the
debate about the extent to which neural networks can
capture the regularities of the linguistic environment,
as we will discuss in chapter 10. Generalization has
also been a major focus in the study of machine learn-
ing (e.g., Weigend, Rumelhart, & Huberman, 1991;
Wolpert, 1996b, 1996a; Vapnik & Chervonenkis, 1971).
In chapter 3, we briefly discussed how networks typ-
ically generalize in the context of distributed represen-
tations. The central idea is that if a network forms dis-
tributed internal representations that encode the compo-
sitional features of the environment in a combinatorial
fashion, then novel stimuli can be processed success-
fully by activating the appropriate novel combination
of representational (hidden) units. Although the com-
bination is novel, the constituent features are familiar
and have been trained to produce or influence appro-
priate outputs, such that the novel combination of fea-
tures should also produce a reasonable result. In the
domain of reading, a network can pronounce nonwords

correctly because it represents the pronunciation con-
sequences of each letter using different units that can
easily be recombined in novel ways for nonwords (as
we will see in chapter 10).
We will see in the next exploration that the basic
GeneRec network, as derived in chapter 5, does not gen-
eralize very well. Interactive networks like GeneRec
are dynamic systems, with activation dynamics (e.g.,
attractors as discussed in chapter 3) that can interfere
with the ability to form novel combinatorial represen-
tations. The units interact with each other too much to
retain the kind of independence necessary for novel re-
combination (O'Reilly, in press, 1996b).
A solution to these problems of too much interactiv-
ity is more systematic attractor structures, for example
articulated or componential attractors (Noelle & Cot-
trell, 1996; Plaut & McClelland, 1993). Purely error-
driven GeneRec does not produce such clean, system-
atic representations, and therefore the attractor dynam-
ics significantly interfere with generalization. As dis-
cussed previously, units in an error-driven network tend
to be lazy, and just make the minimal contribution nec-
essary to solve the problem. We will see that the ad-
dition of Hebbian learning and kWTA inhibitory com-
petition improves generalization by placing important
constraints on learning and the development of system-
atic representations.