Is Occam’s Razor Dulling? How to Sharpen Your Search for Truth

The ancient Greek astronomer Ptolemy believed that the Earth was the center of the solar system. As new observations arose, each contradicted his existing model, necessitating adjustments until Nicolaus Copernicus introduced a revolutionary concept: the Sun as the center of our solar system. This shift initiated a scientific revolution that transformed our understanding of the universe.

Simplistic theories have frequently given way to more robust explanations throughout scientific history. For example, as knowledge advanced, the concept of special relativity replaced the outdated idea of the luminous ether, and the theory of continental drift explained geological features better than the outdated land bridge theory. This mirrors the principle of Ockham’s razor, attributed to the 14th-century philosopher William of Ockham, which posits that the simplest explanation is often the correct one.

But what if scientific advancement doesn’t always favor simplicity? What if the complexities of the universe provide deeper insights? To uncover hidden structures, we may need to shift our perspective.

Cognitive scientist and philosopher Marina Dubova and researchers from the Santa Fe Institute propose that Ockham’s razor could limit our approach to science. Using computer simulations and experimental data from the microscopic world, they apply psychology principles to examine scientists’ methodologies. Dubova suggests that many foundational beliefs we hold about searching for truth may not be as solid as we once thought.

As scientific automation increases, understanding these risks could be essential for the development of “AI scientists.” New Scientist interviewed Dubova to discuss the implications of integrating traditional concepts into modern scientific practices and how we can enhance our learning by maximizing real-world engagement.

Thomas Luton: What is Occam’s Razor Principle? How do scientists utilize it?

Marina Dubova: Occam’s Razor favors simpler explanations. In multiple scientific disciplines, we are often advised to begin with the simplest theory. When anomalies appear, we can introduce additional variables or mechanisms, but starting simple is key. Scientists employ this in various ways, often preferring explanations that minimize assumptions, or selecting theories with fewer causal mechanisms. However, certain situations may call for more defined predictions, leading to less flexible approaches.

Is this simplistic approach unique to science, or do we all engage in it?

Research shows that when individuals explain events, they often opt for broad, simple explanations. Psychologist Tania Lombrozo found that participants typically preferred explanations with fewer causes. For instance, when diagnosing an alien with two symptoms, they leaned toward a single disease explanation over two separate diseases, even if the latter was statistically more probable.

Is there empirical evidence supporting whether Occam’s Razor fosters scientific progress?

In my studies, I used computational models where AI agents developed ground truth representations based on limited datasets. Some agents formulated theories with minimal variables, while others delved into complexity, yielding far more elaborate descriptions. Interestingly, it became apparent that agents favoring complexity could predict new data as effectively, or even better, than their parsimonious counterparts, prompting a reevaluation of our assumptions regarding scientific inquiry.

So, it’s not just Occam’s razor that leads us astray?

Indeed, another pervasive guideline dictates that experiments should base their hypotheses on established theories. Whether assessing extraterrestrial life or exploring human memory, establishing a theory is essential for conducting theory-driven experiments. The legendary solar eclipse expedition by Arthur Eddington in 1919 exemplified this, specifically testing general relativity’s predictions regarding starlight bending due to gravity.

Is the pursuit of reason in science a misguided approach?

Similar computational models can simulate agents employing strategies to disprove theories or resolve disputes through carefully selected experiments. One approach is to conduct affirming experiments, highlighting a tendency towards confirmation bias. Alternatively, agents may engage in exploratory methods, either through random experimentation or introducing novel experiments.

Who formulates the best scientific theories?

Agents employing innovative or random exploratory strategies produced superior theories about underlying truths. This surprising outcome led us to conduct additional experiments to validate our findings.

Have you observed this behavior among real scientists?

Yes. We engaged neuroscientists in evaluating brain imaging to identify causal structures in toy brain models. Although they successfully achieved much, certain neuroscientists struggled to adapt their prior beliefs, insisting that one brain area must correspond to one function. In reality, our model featured regions controlling multiple abilities, challenging the simplistic notion that each area has a singular responsibility. This parallels how real scientists shape their experimental approaches based on preconceived hypotheses.

What insights should scientists glean from your experiment?

Scientific institutions historically resist nurturing expansive inquiry. Recognizing how our concepts shape decisions is crucial in preventing us from understanding reality in progressive ways. For instance, established theories like general relativity or the periodic table can facilitate scientific progress yet simultaneously limit exploration.

Does science naturally evolve, overturning old theories? Isn’t that sufficient?

While scientific revolutions reshape our methods and perspectives significantly, does it necessitate decades or centuries to challenge these ideas? More flexible exploration can expedite discovery.

Can you offer examples of how oversimplified explanations misguide science?

In neuroscience, the shift to viewing the brain as a complex network highlights how early theories of isolated functions have become obsolete. Genetics also reflects this trend, moving beyond the assumption that one gene governs a trait; contemporary understanding reveals that most traits arise from various genes interacting within environmental contexts.

Could simpler models have provided a useful starting point for complex phenomena?

Humans possess cognitive constraints that limit our ability to formulate intricate, high-dimensional models that could encompass all facets of study. This frugality may have been a necessity. Today, artificial intelligence aids us in investigating more nuanced and comprehensive explanations.

This shift is evident in the field of statistical learning, which investigates how systems glean insights from data. A recent finding on “double descent of generalization” has revealed that larger models, previously thought to perform worse, actually demonstrate improved performance, reaffirming the value of complexity in broader scenarios.

How does this influence the training of AI scientists?

I believe we must critically assess which components of the scientific method deserve retention and which warrant reconsideration. Rather than simply adhering to traditional methods due to cognitive constraints, we should develop AI that is equipped to recognize and address inherent biases to avoid perpetuating past pitfalls.

What revelations do your experiments illuminate about the essence of science?

The core goal of science lies in understanding reality and uncovering knowledge about the world. Given the myriad of angles through which phenomena can be examined, a diverse perspective is paramount. Philosopher Haseok Chan from Cambridge University posits that maximizing our contact with reality is fundamental. Science embodies a tactile pursuit of knowledge, reminiscent of the philosophical rigor described by Masvita Chirimuta at the University of Edinburgh, who underscores the significance of interactive exploration.