Summary: Decision-making is often shaped by trial and error, yet many traditional models assume people and animals always act optimally based on past outcomes. A new study uses very small, interpretable artificial neural networks to reveal how individuals actually make choices—highlighting systematic, but frequently suboptimal, strategies used in real settings.
These compact models predict individual decisions more accurately than many classical theories by capturing the imperfect, idiosyncratic behaviors that occur in practice. The findings offer a new way to map cognitive strategies and could influence how we design behavioral and mental health interventions.
Key Facts:
- Realistic insights: Small AI models expose decision-making strategies that are structured yet often suboptimal.
- Individual differences: The models forecast each subject’s behavior more reliably than frameworks that assume optimal choices.
- Practical implications: Understanding diverse decision strategies may inform personalized approaches in mental health, education, and behavioral science.
Source: NYU
Researchers have long studied how humans and animals learn from trial and error, using recent experience to guide future choices. Many conventional frameworks—like reinforcement learning and Bayesian inference—describe useful principles for adaptive behavior, but their emphasis on optimal performance can miss the messy realities of biological decision-making.
A new study introduces an alternative: tiny recurrent neural networks that are small enough to interpret yet capable of capturing complex choice behavior. By fitting these models to behavioral data, the authors were able to uncover the algorithms individuals use to learn and decide—even when those algorithms do not maximize rewards in the idealized sense.
“Rather than prescribing how brains should learn, we sought to discover how individual brains actually learn,” says Marcelo Mattar, assistant professor in New York University’s Department of Psychology and a coauthor of the study published in Nature. The approach treats model fitting like detective work, using compact neural networks to reveal consistent decision strategies that prior models had overlooked.
Because the networks are very small—often just one to four units—the researchers can analyze their dynamics with mathematical tools and dynamical-systems methods. This interpretability is a key advantage over large, black-box AI models, which are often excellent at prediction but hard to summarize succinctly in mechanistic terms.
Ji-An Li, a doctoral student in the Neurosciences Graduate Program at UC San Diego, notes that the simplicity of these networks enables clearer explanations of why a model makes a particular prediction. Marcus Benna, assistant professor of neurobiology at UC San Diego, adds that training small networks and analyzing their internal dynamics allows the team to translate complex learning behavior into more readily understandable strategies.
Across a variety of laboratory reward-learning tasks and species—including humans, non-human primates, and laboratory rats—the tiny networks often outperformed classical cognitive models in predicting choices. In many cases their predictive accuracy matched that of much larger neural networks used in commercial AI, while remaining far more interpretable.
Importantly, these models regularly captured suboptimal choices made by subjects, reflecting the “real-world” mixture of heuristics, biases, and limited information processing that shape behavior. Instead of forcing data into an optimality framework, the small networks reveal the diverse strategies individuals deploy when learning from experience.
The method also allows estimation of the effective dimensionality of behavior and provides insight into the algorithms that meta-reinforcement-learning agents acquire, creating a bridge between cognitive theory and modern AI research.
“Just as recognizing individual biological differences has transformed medicine, mapping individual differences in decision-making strategies could reshape approaches to mental health, education, and cognitive assessment,” Mattar concludes. By offering a systematic, interpretable framework for discovering cognitive strategies, the study supplies tools for both basic science and applied interventions.
Funding: The research received support from grants by the National Science Foundation (CNS-1730158, ACI-1540112, ACI-1541349, OAC-1826967, OAC-2112167, CNS-2100237, CNS-2120019), the Kavli Institute for Brain and Mind, the University of California Office of the President, and UC San Diego’s California Institute for Telecommunications and Information Technology/Qualcomm Institute.
About this AI and decision-making research news
Author: James Devitt
Source: NYU
Contact: James Devitt – NYU
Image: Image credited to Neuroscience News
Original Research: Open access.
Title: “Discovering cognitive strategies with tiny recurrent neural networks” by Marcelo Mattar et al., published in Nature.
Abstract
Discovering cognitive strategies with tiny recurrent neural networks
Understanding how animals and humans learn from experience to make adaptive decisions is a core aim of neuroscience and psychology. Normative frameworks such as Bayesian inference and reinforcement learning clarify general principles of adaptive behavior, but their simplicity and emphasis on optimality can limit their ability to capture realistic biological behavior. This often prompts iterative, handcrafted model adjustments that introduce researcher bias.
The study presents a novel modeling pipeline that leverages recurrent neural networks to discover cognitive algorithms underlying decision-making. Networks with as few as one to four units frequently outperform standard cognitive models and match larger neural networks when predicting individual choices across six established reward-learning tasks. Crucially, these small models can be interpreted using dynamical-systems tools, allowing a unified comparison of cognitive models and revealing fine-grained mechanisms of choice behavior.
The approach estimates behavioral dimensionality and sheds light on algorithms learned by meta-reinforcement-learning agents, providing a principled, interpretable foundation for studying both healthy and dysfunctional cognition.