Page 292
While it may be true that normative discourse cannot be replaced without remainder by descriptive
discourse, it would be a distortion to represent this as the aim of those who would naturalize
epistemology. The aim is rather to enlighten our normative endeavors by reconstructing them within a
more adequate conception of what cognitive activity consists in, and thus to free ourselves from the
burden of factual misconceptions and tunnel vision. It is only the autonomy of epistemology that must
be denied.
Autonomy must be denied because normative issues are never independent of factual matters. This is
easily seen for our judgments of instrumental value, as these always depend on factual premises about
causal sufficiencies and dependencies. But it is also true of our most basic normative concepts and our
judgments of intrinsic value, for these have factual presuppositions as well. We speak of justification,
but we think of it as a feature of belief, and whether or not there are any beliefs and what properties they
have is a robustly factual matter. We speak of rationality, but we think of it as a feature of thinkers, and
it is a substantive factual matter what thinkers are and what cognitive kinematics they harbor. Normative
concepts and normative convictions are thus always hostage to some background factual
presuppositions, and these can always prove to be superficial, confused, or just plain wrong. If they are,
then we may have to rethink whatever normative framework has been erected upon them. The lesson of
the preceding pages is that the time for this has already come.

Page 326
2.2.1 Learning
If a connectionist model is intended to learn, there will be processes that determine the weights of the
connections among its units as a function of the character of its training. Typically in a connectionist
machine (such as a Boltzmann Machine) the weights among connections are adjusted until the system's
behavior comes to model the statistical properties of its inputs. In the limit, the stochastic relations
among machine states recapitulate the stochastic relations among the environmental events that they
represent.
This should bring to mind the old associationist principle that the strength of association between
"ideas" is a function of the frequency with which they are paired "in experience" and the learning-
theoretic idea that the strength of a stimulus-response connection is a function of the frequency with
which the response is rewarded in the presence of the stimulus. But though connectionists, like other
associationists, are committed to learning processes that model statistical properties of inputs and
outputs, the simple mechanisms based on co-occurrence statistics that were the hallmarks of old-
fashioned associationism have been augmented in connectionist models by a number of technical
devices. (Hence the 'new' in 'new connectionism'). For example, some of the earlier limitations of
associative mechanisms are overcome by allowing the network to contain "hidden" units (or aggregates)
that are not directly connected to the environment, and whose purpose is, in effect, to detect statistical
patterns in the activity of the ''visible" units including, perhaps, patterns that are more abstract or more
"global" than the ones that could be detected by old-fashioned perceptrons.
In short, sophisticated versions of the associative principles for weight setting are on offer in the
connectionist literature. The point of present concern, however, is what all versions of these principles
have in common with one another and with older kinds of associationism: namely, that these processes
are allfiequency-sensitive. To return to the example discussed above: if a connectionist learning machine
converges on a state where it is prepared to infer A from A & B (that is, to a state in which, when the 'A
& B' node is excited, it tends to settle into a state in which the 'A' node is excited), the convergence will
typically be caused by statistical properties of the machine's training experience (for instance, by
correlations between firings of the 'A & B' node and firings of the 'A' node, or by correlations of the
firings of both with some feedback signal). Like traditional associationism, connectionism treats
learning as basically a sort of statistical modeling.