Serial, parallel and hierarchical decision making in primates

Ariel Zylberberg, Jeannette A. M. Lorteije, Brian G. Ouellette, Chris I. De Zeeuw, Mariano Sigman, Pieter R. RoelfsemaeLife 2017 · DOI

TL;DR

This paper revisits the debate on how primates make hierarchical decisions, arguing against "flat" models where all options are considered simultaneously. By reanalyzing existing data and presenting new experiments, the authors demonstrate that hierarchical decision-making is a flexible process that can switch between parallel and serial strategies, a capability flat models lack.

What Problem Does This Paper Address?

The way our brains make decisions, especially complex ones involving multiple steps, is a fundamental question in neuroscience. Traditionally, decision-making research focused on simple, single choices. However, real-world problems often require breaking down a larger decision into a series of smaller, interconnected sub-decisions. A key question is whether these sub-decisions are processed in parallel (all at once) or serially (one after another). Previous work suggested primates might use a hierarchical, parallel approach, but a competing "flat" model proposed that all final options are evaluated simultaneously. This paper aims to definitively resolve this debate by testing the predictions of hierarchical versus flat models against behavioral and neural data.

Methods

The authors employ a multi-pronged approach:

  • Reanalysis of Existing Data: They re-examine behavioral and neural data from a previous study (Lorteije et al., 2015) that used a hierarchical decision-making task in monkeys. This involves fitting and comparing different computational models to the data.
  • New Behavioral Experiments: They conduct experiments with human volunteers on similar hierarchical decision-making tasks, but with variations in task complexity (two vs. three levels of decisions).
  • Computational Modeling: They simulate and test "flat" models, specifically addressing criticisms and alternative implementations proposed by other researchers (Hyafil and Moreno-Bote). They also compare these to predictions of hierarchical models.
  • Neurophysiological Recordings: They analyze neural activity (specifically in visual cortex areas V1 and V4) from the monkey experiments to see how different levels of evidence influence neuronal responses.

Datasets

The paper primarily uses data from two sources:

  • Monkey Behavioral and Neural Data: This is the core dataset from Lorteije et al. (2015), involving monkeys performing a hierarchical decision-making task with visual stimuli.
  • Human Behavioral Data: New data collected from three human subjects performing similar hierarchical tasks with varying levels of complexity.

Key Insights and Ideas

  • Flat Models Fall Short: The paper strongly argues that "flat" models, which assume a single competition between all final options, cannot adequately explain the observed behavioral and neural data, particularly the interactions between decision levels.
  • Hierarchical Decision-Making is Flexible: The authors present compelling evidence that hierarchical decision-making is not rigidly parallel. Instead, it can adapt and switch between parallel and serial processing depending on task structure and complexity.
  • Task Structure Matters: The "exclusivity-or" structure of some decision branches (where choosing one option precludes another) forces a more serial processing, which flat models struggle to accommodate.
  • Limited Capacity for Parallelism: A key finding from the human experiments is that there appears to be a limit to the number of decisions that can be processed in parallel. When this limit is exceeded (e.g., by adding a third decision level), the process shifts towards serial processing.
  • Neural Evidence Supports Hierarchy: The neurophysiological data shows distinct patterns of influence between decision levels that are more consistent with a hierarchical structure than a flat one.

Key Results

  • Behavioral Data: Reanalysis of monkey data and new human data show that flat models fail to predict the observed psychophysical kernels and the interaction between decision levels. Specifically, the influence of L1 difficulty on L2 accuracy, predicted by flat models, was not consistently observed.
  • Neural Data: In monkeys, evidence from L1 selectively modulated activity related to L2 branches in V4, with effects that were not uniform across all branches, contradicting the predictions of simple flat models with a single inhibitory source.
  • Human Experiments: In humans, a two-level hierarchical task showed parallel evidence accumulation. However, adding a third level (L3) induced a more serial processing, with evidence for L1 accumulating before L2, and L2 before L3. This suggests a limited capacity for parallel processing.

Limitations and Future Work

  • Defining the Parallelism Limit: The exact capacity for parallel decision-making in humans remains to be precisely determined, though the study suggests it lies between 3 and 7 decisions.
  • Mechanisms of Interaction: While the paper argues for hierarchical processing and shows how local selection signals interact, further research is needed to fully elucidate the specific neural mechanisms underlying these interactions.
  • Role of Practice: The paper notes that prolonged training might lead to more efficient decision-making, but their experiments with highly trained subjects still showed serial processing in complex tasks, suggesting practice doesn't entirely overcome the inherent serial nature in certain conditions.
  • Future Directions: The authors suggest future studies could record from neurons involved in persistent activity (e.g., in LIP or FEF) using variants of these tasks. They also propose investigating the neural mechanisms that improve interactions between successive decisions with increased proficiency.

Summary updated

Abstract

The study of decision-making has mainly focused on isolated decisions where choices are associated with motor actions. However, problem-solving often involves considering a hierarchy of sub-decisions. In a recent study (Lorteije et al. 2015), we reported behavioral and neuronal evidence for hierarchical decision making in a task with a small decision tree. We observed a first phase of parallel evidence integration for multiple sub-decisions, followed by a phase in which the overall strategy formed. It has been suggested that a 'flat' competition between the ultimate motor actions might also explain these results. A reanalysis of the data does not support the critical predictions of flat models. We also examined the time-course of decision making in other, related tasks and report conditions where evidence integration for successive decisions is decoupled, which excludes flat models. We conclude that the flexibility of decision-making implies that the strategies are genuinely hierarchical.
BibTeX
@inproceedings{zylberberg2017serial,
  author = {Ariel Zylberberg and Jeannette A. M. Lorteije and Brian G. Ouellette and Chris I. De Zeeuw and Mariano Sigman and Pieter R. Roelfsema},
  title = {{Serial, parallel and hierarchical decision making in primates}},
  booktitle = {eLife},
  year = {2017},
  doi = {10.7554/elife.17331},
}
Figure