Medieval friar William of Ockham posited a famous idea: always pick the simplest explanation. Often referred to as the parsimony principle, "Ockham's razor" has shaped scientific decisions for centuries.
But lately, incredibly complex AI models have begun outperforming their simpler counterparts. Consider AlphaFold for predicting protein structures, or ChatGPT and its competitors for generating humanlike text.
A new paper in PNAS argues that by relying too much on parsimony in modeling, scientists make mistakes and miss opportunities.
First author and SFI Complexity Postdoctoral Fellow Marina Dubova says overreliance on parsimony is historical.
"Scientists need a tool to guide how they build models of the world. Parsimony was historically adopted as an easy tool to use. Since then, it's not been questioned enough. Educational programs teach parsimony as a key principle in scientific theory and model building. Most research tries to justify why parsimony is good, but those justifications haven't stood the test of time," she says.
Dubova recently ran a computational simulation showing that random experiments generated better models than did scientific experiments chosen based on previous assumptions.
Now Dubova, a cognitive scientist, is probing one of the biggest scientific assumptions of all: avoiding complex models.
"Relying on parsimony alone as our guiding principle limits what we can learn about the world and potentially drives us in wrong directions," says Dubova. "Parsimony and complexity are complementary tools. Scientists need to use evidence, judgment, and context-specific demands to determine whether a more parsimonious or complex model suits their research goals."
Dubova and co-authors discuss findings that suggest misapplied parsimony can make models biased and lead to bad predictions. For example, simple models for interpreting live brain scans often read periodic back-and-forth patterns when, in fact, brain activity is changing slowly over time. Leaving out key characteristics (like patient age) from a model evaluating untested new drugs could lead to poor predictions of who will and won't respond well.
By contrast, complex models can be more flexible and accurate, as new approaches in climate change research have shown. Often in science, each lab develops its own model for making predictions about the phenomenon of interest, and the field eventually converges on the most parsimonious model that best fits the data. However, climate scientists have found that when they combine dozens of sometimes contradictory models from different labs into one ensemble, climate forecasts get better at predicting actual real-world phenomena.
"Even when these climate models are incompatible, scientists decide to employ them all because they know each one is capturing some aspect of the world. The literature suggests that using them together helps us better predict the reality around us," she says. "Could this approach inspire completely new understandings of what climate is, without us as scientists trying to impose our preference for just one simple explanation?"
Dubova hopes that the paper will kickstart new research into when scientific modelers should choose parsimony or complexity.
