The ancient Greek astronomer Ptolemy conceived that the planets and the sun revolved around the Earth, based on the limited knowledge of his era. Every observation that contradicted this notion required a slight adjustment to the theory until Nicolaus Copernicus radically redefined the understanding of our solar system. He proposed that all planets orbit the sun, igniting a scientific revolution that transformed our comprehension of the universe.

Throughout history, simpler theories have often evolved into accepted knowledge. Special relativity emerged victoriously over the concept of the luminous ether, and continental drift gained clarity through shared fossils found across distant continents, rather than imagined land bridges. This aligns with Ockham’s razor, attributed to the 14th-century monk William of Ockham, advocating for the simplest explanation that aligns with the facts.

But what if scientific advancement doesn’t always adhere to simplicity? What if starting with complexity, instead, broadens our understanding and reveals concealed structures?

Cognitive scientist and philosopher Marina Dubova from the Santa Fe Institute in New Mexico suggests that Ockham’s razor is merely one of several guidelines that may cloud our quest for genuine reality. Through computer simulations and microscopic explorations, Dubova applies principles of psychology and cognition to challenge existing scientific paradigms, revealing the fluidity of our assumptions about truth-seeking.

As we advance toward a future where automation in science is likely, these insights could prove instrumental in shaping the development of “AI scientists.” New Scientist reached out to Dubova to discuss the hazards of integrating antiquated concepts into contemporary science, how maximizing real-world interactions can enhance learning speed, and the fundamental implications this holds for the essence of science itself.

Thomas Luton: What is the Occam’s Razor Principle, and how do scientists implement it?

Marina Dubova: Occam’s Razor inherently promotes simplicity in explanations. Students across diverse scientific fields, myself included, are often instructed to commence with the most straightforward theory. If discrepancies arise in the data, additional variables can be introduced, but the initial approach should always be the simplest. Scientists apply this in varied contexts, favoring explanations with minimal assumptions or fewer causative mechanisms. Sometimes, less adaptable explanations are preferred for making specific predictions.

Is this simplistic approach unique to science, or do we all tend to do this?

Evidence suggests that when prompted to clarify various phenomena, humans gravitate towards broad, simplistic explanations. Research conducted by psychologist Tania Lombroso at Princeton University revealed that individuals often prefer explanations invoking fewer causes that adequately account for most data. For instance, participants asked to diagnose an alien showcasing two symptoms preferred a singular ailment encompassing both symptoms, rather than suggesting two different diseases for each symptom, even when the latter presented as more likely.

Is there empirical evidence supporting the notion that Occam’s Razor promotes scientific progress?

My research examined these theoretical concepts using computational models. In this study, an AI agent formed a foundational representation based on limited datasets, generating theories with minimal variables, while others constructed more intricate frameworks. For example, an agent with access to three significant variables might create an explanation with a thousand. Surprisingly, agents prioritizing complexity sometimes performed predictions as well as—or even better than—those favoring simplicity, challenging many entrenched assumptions held by scientists regarding our understanding of the world.

You implied that there are multiple theories that can mislead us, not just Occam’s Razor?

Indeed, another common guideline suggests that experiments should arise from existent theories. In studying extraterrestrial life or human memory phenomena, researchers often formulate a theory to conduct theory-based experiments. An example is Arthur Stanley Eddington’s 1919 solar eclipse expedition, specifically tailored to verify general relativity’s prediction of starlight bending due to the Sun’s gravity, forcing a choice between conflicting hypotheses of Einstein and Newton.

Is the pursuit of reason in science misguided?

Similar computational models create agents that resolve discrepancies using targeted experiments. Another approach involves testing theories for validation, which can introduce confirmation bias. Further exploratory strategies include randomly selecting novel experiments.

Which theoretical approach yielded the most effective theories?

Agents employing exploratory strategies, whether randomness or novelty-driven, produced the most accurate theories about the underlying truth. Our surprising findings prompted additional experiments to revalidate our results.

Have you witnessed similar behaviors in actual scientists?

Yes. We engaged neuroscientists to utilize brain imaging and lesions in deciphering the underlying causal structures of a model brain. Although they succeeded in their task, some showed difficulty in adapting their preconceived notions. For instance, when regions governing multiple functions were evident, some neuroscientists insisted on the existence of distinct areas for each ability, reflecting the tendency to maintain rigid hypotheses even in the face of contrary evidence.

What insights should scientists draw from your research?

Current scientific institutions often inhibit inquiry. It’s crucial to acknowledge that our theories can shape decision-making and limit our exploration of reality. This range of theories—from general relativity to chemical structures—guides and enhances our progress but can also obstruct our understanding of objective reality.

Are scientific revolutions sufficient for overcoming erroneous concepts?

Though scientific revolutions yield significant breakthroughs, will it take countless years to overturn established notions? A more exploratory approach may accelerate discoveries.

Can you provide examples of how simplistic theories led to misconceptions?

Consider advancements in neuroscience, which has shifted to viewing the brain as an integrated network rather than a collection of isolated regions tasked with specific functions. A similar shift in genetics challenges the notion of single-gene traits, revealing a more complex interplay of multiple genes and environmental influences in establishing traits.

One might argue that simpler initial ideas have at least served as a starting point for understanding complex phenomena.

Given our cognitive limitations, such simplifications are somewhat inevitable. However, AI can initiate explorations into higher-dimensional complexities that were previously inconceivable.

This phenomenon is evident in statistical learning, a field that analyzes data interactions. A noteworthy discovery termed “double descent of generalization” illustrates that larger models might have unexpected performance advantages. Historically, it was believed that minimizing environmental details was essential for optimal model performance. Recent findings indicate a reduction in error rates as models grow more complex, suggesting that they excel in predicting unseen data by emphasizing broader representations over mere memorization.

Does this influence the training of AI scientists?

We need robust discussions on which components of the scientific method merit preservation and which need reevaluation. Was our previous scientific approach our best effort given cognitive constraints? Understanding this is critical as we transition toward automating scientific processes; without it, established biases and blind spots may scale up.

What do your findings indicate about the nature of science?

The goal of science is to comprehend reality and expand our knowledge of the world. However, numerous perspectives must be employed to adequately investigate phenomena. Philosopher Haseok Chan from Cambridge University describes this as maximizing our contact with reality. Scientific inquiry often resembles tactile exploration rather than purely visual observation. Engaging with various elements of reality allows us to extract unique insights and understanding.