CNS*2026 Halifax

Introduction
Humans and animals deploy diverse strategies within their ecological environments. Much of neuroscience has focused on normative frameworks in which agents optimize decisions — Bayesian inference, speed–accuracy tradeoffs, gradient-based learning [1,2]. Yet real agents frequently rely on heuristics specifically adapted to their ecology: cognitive shortcuts that save time and resources, and can outperform optimal strategies in complex environments [3]. Despite their influence in psychology and behavioral economics as fast, automatic "System 1" processes, heuristics remain underrepresented in neuroscience. How neural systems implement and switch between heuristics and deliberative strategies remains a central challenge.

Methods
We study heuristics in perceptual decision-making using a discrimination task where one “over-represented” stimulus evokes stronger population activity. Monkeys were shown to solve this task by relying on a simple heuristic: summing activity to detect the over-represented stimulus, instead of optimally integrating activity weighted by task relevance [3]. We implement a model neural network trained with either gradient descent or Oja learning rule under over-represented and equally-represented conditions. We then assess, as in [3], whether optimal readout or heuristic strategies were implemented using: (1) accuracy imbalance between stimuli under inactivation; (2) choice probability correlating single-neuron activity with stimulus choice.

Results
While networks trained with gradient descent learn optimal strategies, networks trained with the Oja rule reproduce empirical signatures of heuristics when the activity is uncentered. Using Oja, networks learn to extract the first principal component of population activity as readout weights [4]. When activity is uncentered, the mean dominates, driving Oja toward a constant readout weight vector "summation" solution. When stimuli become equally-represented, this solution fails and the network centers activity through slow integration of mean activity, yielding a non-heuristic solution consistent with theoretical and experimental predictions. Our framework provides a mechanistic model of switching between heuristic and optimal strategies.

Discussion
Understanding how the brain switches between fast heuristics (System 1) and deliberate cognition (System 2) has broad implications for psychology, neuroscience, and economics [5]. Our framework suggests heuristics are not mere shortcuts but ecologically rational strategies — tuned to environmental statistics and implemented through simple neural computations. This reframing has consequences for how we study decision-making across species. Beyond basic science, these insights can inform neuroAI systems that integrate rapid heuristic strategies with precise reasoning, offering a path toward more generalizable, energy-efficient models.

References

1. Körding, K. P., & Wolpert, D. M. (2004). Bayesian integration in sensorimotor learning. Nature, 427(6971), 244–247. https://doi.org/10.1038/nature02169
   
2. Richards, B. A., & Kording, K. P. (2023). The study of plasticity has always been about gradients. The Journal of Physiology, 601(15), 3141–3149. https://doi.org/10.1113/JP282747
   
3. Laamerad, P., Krause, M. R., Guitton, D., & Pack, C. C. (2025). Inactivation of primate cortex reveals inductive biases in visual learning. Current Biology, 35(19), 4699–4713.e6. https://doi.org/10.1016/j.cub.2025.08.027
   
4. Oja, E. (1982). A simplified neuron model as a principal component analyzer. Journal of Mathematical Biology, 15(3), 267–273. https://doi.org/10.1007/BF00275687
   
5. Kahneman, D. (2011). Thinking, fast and slow. Farrar, Straus and Giroux.
   

Acknowledgement
This work was supported by IVADO Projet Exploratoire (Explo24CO-3750823649).