Page 351 - Deep Learning

334 Conversion

useful. Veridicality, confidence, desirability, and usefulness are distinct types
of judgments. There might be yet other dimensions along which propositions
can be judged but these suffice for present purposes.

The Structure of Belief Systems

We have no method for counting beliefs, but we can estimate the size of a
person’s belief base from the size of his vocabulary. The average educated per-
son in a Western society knows approximately 50,000 words. Some words are
8
synonyms, but many words have more than one meaning; suppose that these
factors approximately cancel each other. A person’s stock of concepts is then
likewise to be estimated at 50,000. Most concepts participate in more than one
belief. Consider the concept dog: Common beliefs about dogs include dogs are
friendly; dogs like biscuits; dogs have four legs; and so on. The total number of
beliefs is likely to be an order of magnitude greater than the number of con-
cepts. According to this estimate, a person’s belief base contains at least half
a million beliefs. A theory of belief formation should explain how this vast
system is structured and how it grows over time.
The standard network model of long-term memory invites the view that
a belief base is a set of belief nodes linked via binary relations (implies, contra-
dicts, special case of, etc.). This view implies that we can only talk about the
9
truth value of a belief system or intuitive theory as shorthand for the truth
values associated with its constituent beliefs. This is a misleading side effect
of the incompleteness of the network model. Within that model, there is no
natural way to represent an area within a network – a subnet – as forming a
separate unit, and the network concept provides no tools for assigning belief
parameters to such a subnet. But this is a deficiency in the network model of
memory rather than a property of memory itself.
Beliefs are grouped in memory by topic or theme into semi-independent
belief systems. A belief system is locally bounded in the sense that activation
of one belief in a belief system leads with high probability to the activation
of the other beliefs in that system (but not necessarily beliefs in neighboring
systems). Declarative knowledge is completely interconnected – there is some
sequence of links from any belief to any other belief – but it is also structured
in the sense that a belief system is a bounded, local and semi-independent
area within the overall network. Exactly how the cognitive architecture man-
ages to have it both ways, to maintain local boundaries within the knowledge
network without losing complete interconnectedness, is not known.