188
9 Networked Minds
while one would expect just the opposite! In other words, when complex problems
must be solved, diversity wins, not the best. This seems really counter-intuitive:
nobody is right, but the average of answers, none of which is good enough, gives
the most accurate picture! The reason for this is subtle and extremely important for
collective wisdom. Although each of the top teams almost achieved a 10% improvement over the original algorithm, each used different methods that were able to find
different patterns in the data. However, no single algorithm could find them all, and
“overfitting” (i.e. fitting random patterns that did not have a meaning) was probably
a problem, too. By averaging the predictions, each algorithm contributed the knowledge it was specialized to find, while the errors of each algorithm were neutralized
by the others.
Even more surprisingly, when more accurate predictions were given a higher
weight, it didn’t improve the accuracy of the average prediction. Researchers have
argued that this is because weighting more successful algorithms more highly only
works if at least one algorithm is correct. In the case of the Netflix challenge,
however, no single algorithm was perfect, and an equally weighted combination was
better than any individual algorithm and also better than an individually weighted
average that took the relative rank of each algorithm into account.
245 This is probably
the best argument for equal votes—in this case, however, equal votes for different
approaches to problem-solving! Therefore, if we want to maximize collective intelligence, compared to our current mode of decision-making, we would have to take
on board more knowledge of minorities. This would give rise to a “democracy of
ideas”, which would allow us to make better use of ideas and innovations, to the
benefit of our society.
9.13 Future Decision-Making Institutions
To conclude, when complex tasks are to be solved, specialization and diversity are
key. So, what decision-making procedures and institutions do we need to maximize
collective intelligence? Answering this question is of major importance to cope with
the increasing difficulty of the challenges posed by our complex globalized world.
We are in the awkward situation that no single person or computer system alone can
fully grasp this complexity, and that we need to rely on inputs from others. So, for
collective intelligence to work, having a knowledge base of trustable and unbiased
information is essential, i.e. measures against information pollution are important.
20
This requires a sufficient level of transparency and quality-ensuring mechanisms,
and it requires a participatory (“crowd sourcing”) approach to get the best ideas on
board (think, for example, of Wikipedia or GitHub).
20 In fact, to avoid mistakes, the more we are flooded with information the more must we be able to
rely on it, as we have increasingly less time to judge its quality.
9 Networked Minds
while one would expect just the opposite! In other words, when complex problems
must be solved, diversity wins, not the best. This seems really counter-intuitive:
nobody is right, but the average of answers, none of which is good enough, gives
the most accurate picture! The reason for this is subtle and extremely important for
collective wisdom. Although each of the top teams almost achieved a 10% improvement over the original algorithm, each used different methods that were able to find
different patterns in the data. However, no single algorithm could find them all, and
“overfitting” (i.e. fitting random patterns that did not have a meaning) was probably
a problem, too. By averaging the predictions, each algorithm contributed the knowledge it was specialized to find, while the errors of each algorithm were neutralized
by the others.
Even more surprisingly, when more accurate predictions were given a higher
weight, it didn’t improve the accuracy of the average prediction. Researchers have
argued that this is because weighting more successful algorithms more highly only
works if at least one algorithm is correct. In the case of the Netflix challenge,
however, no single algorithm was perfect, and an equally weighted combination was
better than any individual algorithm and also better than an individually weighted
average that took the relative rank of each algorithm into account.
245 This is probably
the best argument for equal votes—in this case, however, equal votes for different
approaches to problem-solving! Therefore, if we want to maximize collective intelligence, compared to our current mode of decision-making, we would have to take
on board more knowledge of minorities. This would give rise to a “democracy of
ideas”, which would allow us to make better use of ideas and innovations, to the
benefit of our society.
9.13 Future Decision-Making Institutions
To conclude, when complex tasks are to be solved, specialization and diversity are
key. So, what decision-making procedures and institutions do we need to maximize
collective intelligence? Answering this question is of major importance to cope with
the increasing difficulty of the challenges posed by our complex globalized world.
We are in the awkward situation that no single person or computer system alone can
fully grasp this complexity, and that we need to rely on inputs from others. So, for
collective intelligence to work, having a knowledge base of trustable and unbiased
information is essential, i.e. measures against information pollution are important.
20
This requires a sufficient level of transparency and quality-ensuring mechanisms,
and it requires a participatory (“crowd sourcing”) approach to get the best ideas on
board (think, for example, of Wikipedia or GitHub).
20 In fact, to avoid mistakes, the more we are flooded with information the more must we be able to
rely on it, as we have increasingly less time to judge its quality.