A recent article suggests that an excessive focus on simplicity in scientific modeling can lead to errors and missed opportunities.
The medieval thinker William of Ockham famously proposed a principle: choose the simplest explanation. Known as “Ockham’s razor,” this concept has influenced scientific reasoning for centuries.
However, highly intricate AI models are now showing better performance than their simpler versions. For instance, AlphaFold excels in predicting protein structures, while ChatGPT and similar models generate text that closely resembles human writing.
A new study published in PNAS claims that an overreliance on simplicity in modeling leads scientists to make errors and overlook potential advancements.
Marina Dubova, the lead author and a Complexity Postdoctoral Fellow at SFI, highlights that this tendency towards simplicity is deeply rooted in scientific history.
“Historically, scientists adopted parsimony as an easy guideline for crafting models of reality. Since then, it has gone largely unchallenged. Educational systems instill parsimony as a fundamental principle in scientific theory and model development. Most research seeks to validate the benefits of parsimony, but those justifications have not always stood up over time,” she explains.
Through a computational simulation, Dubova found that random experimentation often produced better models than those derived from pre-established scientific assumptions.
As a cognitive scientist, Dubova is now questioning a key scientific belief: the avoidance of complex models.
“Relying solely on parsimony restricts our understanding of the world and can lead us astray,” Dubova warns. “Parsimony and complexity should be seen as complementary approaches. Scientists need to assess evidence, use their judgment, and consider the specific context to choose whether a simpler or more complex model is best suited for their research aims.”
Dubova and her co-authors point out that misapplied simplicity can bias models and result in inaccurate predictions. For instance, simplistic models interpreting live brain scan data may misinterpret slow-changing brain activity as rapid oscillations. Additionally, excluding critical variables (like the age of patients) from models analyzing unproven new drugs could result in inaccurate forecasts regarding patient responses.
In contrast, complex models often demonstrate greater flexibility and precision, as shown by recent developments in climate change research. Typically, different laboratories create their distinct models to forecast specific phenomena, eventually leading the field to coalesce around the simplest model that aligns with the data. However, climate scientists have discovered that combining numerous sometimes conflicting models from various labs into a single ensemble can enhance the accuracy of predictions regarding real-world occurrences.
“Even when these climate models contradict each other, scientists utilize them all because each captures different aspects of reality. Research indicates that using these models together improves our predictive capabilities,” she states. “Might this strategy lead to entirely new insights about climate, bypassing the tendency of scientists to gravitate towards a single straightforward explanation?”
Dubova is hopeful that this paper will stimulate further research into how scientific modelers can better decide when to opt for simplicity or complexity.
