Page 332 - Deep Learning

The Formation of Belief 315

relation between theory and data within some logical or mathematical cal-
culus did not attract the new generation. Thomas Kuhn’s The Structure of
Scientific Revolutions, first published in 1962 but better known in its second,
53
1970, edition, broke with the formal approach. Kuhn’s critique of Popper
was blunt: “no process yet disclosed by the historical study of [science] at all
resembles the methodological stereotype of falsification by direct comparison
54
with nature.” This was a declaration that a philosophy of science should be
accountable to data from the history of science and describe how episodes of
theory change unfold.
Using the Copernican conversion from a geocentered to a heliocentered
theory of the planetary system as his main example, Kuhn argued three related
theses. First, in the normal case, scientists do not conduct studies to test their
theories. instead, they use those theories to solve problems, explain phenom-
ena, interpret data or answer questions. Kuhn called this activity puzzle solving;
perhaps his work would have been better received if he had used the less belit-
tling label “problem solving.” Although Kuhn’s main examples are drawn from
physics, it is not difficult to think of examples from other sciences. A chemist
who studies a chemical reaction does not usually intend to verify or falsify the
atomic theory of matter, the periodic table or the theory of the co-valent bond.
He uses those theories to describe how the reaction of interest unfolds. When
researchers agree on what constitutes a worthwhile problem and on the crite-
ria for a successful solution, they are engaged in normal science.
if scientists normally apply rather than test their theories, how and why do
they ever abandon them? Kuhn’s second thesis was that scientists sometimes fail
to solve the problem they set themselves. The phlogiston theory enabled chemists
to solve a good many problems in chemistry, but not all; Jean-Baptiste Lamarck’s
theory of evolution enabled biologists to understand the fitness between a spe-
cies and its environment, but not the geographic distribution of species; and so
on. An unsolved research problem is a cause for dissatisfaction, but the state-
ment I believe this theory is true but I have so far been unable to solve problem
so­and­so does not express a logical contradiction. The problem might be (tem-
porarily) unsolvable for other reasons than that the theory is false, including
limits on the scientist’s creativity or mathematical skills, undetected equipment
failures or lack of data. Kuhn did not emphasize (and his many commentators
tend to overlook) that his reconceptualization of cognitive conflict injected a
pragmatist thread into the study of theory change. The trigger for change is not
falsifying evidence but failure to solve a problem, a very different concept.
sensitivity to historical data led Kuhn to describe theory change as a
protracted affair. An unsolved problem is an anomaly only if it has resisted