Cognitive Scientist Models Human 'Mistakes' as Rational Adaptations
On the Super Data Science podcast, Princeton Professor Tom Griffiths explained his theory that many human cognitive 'mistakes' are actually rational adaptations to real-world constraints. He argues that humans act as "intuitive data scientists," making statistical inferences that seem irrational to psychologists but impressive to computer scientists because they account for computational limitations and incomplete information.
- The theory is formally known as "resource-rational analysis," which models human cognition as the optimal use of limited computational resources. This approach reframes seemingly irrational heuristics and biases as efficient solutions to complex problems under constraints like limited time and processing power. - Griffiths' work often employs Bayesian statistics to model how humans make inferences and learn from limited data, a concept central to modern machine learning. This perspective treats the human mind as a system that excels at inductive reasoning, drawing strong conclusions from sparse information by leveraging prior knowledge, much like Bayesian models in AI. - In the book *Algorithms to Live By*, co-authored by Griffiths, he explores how algorithms used in computer science can be applied to solve human decision-making problems. For instance, the book discusses the "37% Rule" from optimal stopping theory as a strategy for problems like apartment hunting or hiring, demonstrating a direct link between computational algorithms and rational human strategies. - The concept of "bounded rationality," a foundational idea in this work, is increasingly relevant in AI and automated machine learning (AutoML). It acknowledges that perfectly optimal decisions are often computationally intractable, leading to the use of heuristics and satisficing—finding "good enough" solutions—a principle applied in complex systems at companies like Google and Meta. - This research connects directly to the trade-offs seen in large-scale ML systems, such as the balance between accuracy and latency in recommendation engines. A resource-rational approach would suggest that a model that is slightly less accurate but significantly faster may represent a more optimal solution in a production environment, a key consideration in MLOps. - The idea of resource rationality has parallels in reinforcement learning, where an agent must learn to act optimally given constraints on its computational resources and available information. This framework can be used to design more efficient learning algorithms that better mimic the resourcefulness of human intelligence. - For those interested in further reading, Griffiths' papers are often published in top-tier conferences like NeurIPS and journals such as *Behavioral and Brain Sciences*, which are highly regarded by researchers at FAANG companies. His work provides a vocabulary for discussing trade-offs in system design that can be valuable in technical interviews.
