From a brain science perspective, the method just described is unsatisfactory because it is obvious
that complex hypotheses such as crime explanations and scientific claims are not represented in the
brain by single neurons. Another problem is that constraints in figure 4.2 are symmetric, allowing two
elements to mutually constrain each other, so that excitatory and inhibitory links between units are
also symmetric. But in real neural networks it never happens that two neurons both excite or inhibit
each other. Fortunately, it is possible to model the calculation of explanatory coherence in a much
more neurologically realistic way.
First, we represent each element by a population of neurons rather than by a single unit. Second,
we represent a link by a whole complex of links between neurons in multiple populations. At this
level, there is no problem in having symmetrical connections because some of the neurons in one
population excite neurons in the other, while others in the second population excite other neurons in
the first one. The resulting neural networks with thousands of artificial neurons are much larger than
the few dozen units that suffice for modeling the Simpson trial and other legal and scientific cases.
But your brain has approximately one hundred billion neurons to work with, so this scale does not
seem to be a problem. Our computer simulations show that more biologically realistic neural
networks can accomplish the same kind of parallel constraint satisfaction as can the simpler ones that
use one unit for each element.
Thus we are beginning to understand how inference to the best explanation might occur in the brain.
As we saw with perception, inference is the result of the dynamic parallel interaction of neural
patterns, not of serial linguistic steps. Attention to brain mechanisms shows how inference can
nonmysteriously be holistic and multimodal. Later chapters will show how similar neural
mechanisms tie inference with emotions and actions.
Coherence and Truth
Brains know reality through a combination of perception and inference to the best explanation of what
is observed. Such inference attempts to maximize explanatory coherence, which sometimes requires
rejection of what the senses tell us. Rejection of observations occurs both in everyday life, as when a
drunk decides that a double image of a person cannot be right, and in science, as when a researcher
throws out some experimental data that conflict with a well-supported theory. Hence gaining
knowledge is a matter of seeking coherence among many hypotheses and pieces of evidence, not of
starting with some indubitable foundation in sense experience or a priori knowledge and trying to
base everything else on that.
There are no foundations for knowledge, even though the overall reliability of perception justifies
recognizing that the results of observation should have a degree of priority in the maximization of
coherence. Knowledge is not a matter of pure coherence, because observational evidence does get
some priority and provides some constraint on utterly fanciful speculation. Nevertheless, I advocate a
kind of coherentism, the view that beliefs are justified by how well they fit with other beliefs and
with sensory experience. This coherentist view of knowledge meshes well with constructive realism
to provide an answer to the two questions that began this chapter: what is reality, and how do we
know it? Reality consists of objects and their properties that we can learn about through perception
and inference to the best explanation.
In chapter 2, I described how inference to the best explanation in science differs from everyday