Distinct decision processes for 3D and motion stimuli in both humans and monkeys revealed by computational modelling

Revan Rangotis, Sylwia Nowakowska, Peter Dayan, Andrew Parker, Ehsan Kakaei, Akinlolu Akande, Kristine KrugbioRxiv 2024 · DOI

TL;DR

By combining behavioral experiments with advanced computational modeling, researchers have discovered that humans and monkeys use fundamentally different decision-making strategies when judging 3D depth versus simple motion. These distinct "decision pipelines" remain consistent across species and are independent of whether the subject responds with their eyes or their hands, suggesting that the nature of the sensory evidence itself—rather than the motor output—shapes how our brains commit to a choice.

What Problem Does This Paper Address?

In the field of systems neuroscience, area V5/MT is a well-known "hub" for processing visual information. It is essential for both perceiving the direction of moving dots (motion) and understanding the 3D structure of objects (depth). Because both types of information pass through the same neural real estate, many researchers assumed they might share a common "decision path"—a universal way the brain accumulates evidence to reach a verdict.

The authors sought to test this assumption. They asked: Does the brain use the same computational "machinery" to decide if a cylinder is rotating clockwise as it does to decide if dots are moving left? Furthermore, does the way we report that decision (using a hand gesture vs. an eye movement) or the species (human vs. macaque) change the underlying math of the decision process?

Methods

The researchers employed two sophisticated modeling frameworks to "peek" under the hood of behavior:

  1. Hierarchical Drift Diffusion Model (HDDM): This is a standard model for two-choice tasks. It assumes the brain accumulates noisy evidence over time until it hits a "boundary." Key parameters include boundary separation (how much evidence is needed) and drift rate (the quality/speed of the evidence).
  2. GLM-HMM (Generalized Linear Model - Hidden Markov Model): Unlike the DDM, which assumes a fixed strategy, the GLM-HMM allows for "latent states." It detects if a subject is currently "strongly engaged," "biased," or "disengaged," and how they switch between these states over time.
  3. Linear Discriminant Analysis (LDA): A dimensionality reduction technique used to determine which specific model parameters were most responsible for the differences observed between tasks.

Datasets

The study utilized two primary behavioral datasets:

  • Human Dataset: 20 healthy adults who performed four combinations of tasks (3D vs. Motion) and response modes (Eye vs. Hand).
  • Monkey Dataset: 2 rhesus macaques (m133 and m134). These monkeys were highly trained: one performed the motion task with hand responses and the 3D task with eye movements.

The tasks involved Random-Dot Kinematograms (RDK) for motion and Structure-from-Motion (SfM) cylinders for 3D depth.

Key Insights and Ideas

  • Stimulus Trumps Response: The most striking finding was that the type of visual stimulus (3D vs. motion) caused massive shifts in decision parameters, while the mode of response (eye vs. hand) had almost no effect. This suggests that decision-making is "upstream" of the motor system.
  • The "Integration Window" Difference: 3D depth decisions consistently showed lower boundary separation ($a$) than motion decisions. This implies the brain uses a shorter "integration window" for 3D signals, committing to a choice faster once a certain threshold is met, possibly due to internal feedback loops that stabilize 3D percepts more quickly.
  • Cross-Species Unity: Despite millions of years of evolution, humans and monkeys showed remarkably similar shifts in their decision strategies when moving from motion to 3D tasks.
  • Latent Strategies: The GLM-HMM revealed that when judging 3D cylinders, subjects stay in a "strongly engaged" state for the vast majority of the time. In contrast, during motion tasks, they frequently flip between different engaged states, one of which is often heavily biased.

Key Results

  • Boundary Separation: In both species, the boundary ($a$) was significantly lower for the 3D cylinder task than for the RDK motion task.
  • LDA Performance: The LDA showed that boundary separation accounted for 69% of the discriminability between the two tasks, confirming it as the primary "lever" the brain pulls to distinguish these processes.
  • Reaction Times: Monkeys were generally faster but less accurate than humans. Specifically, monkeys were much faster at responding to 3D stimuli with eye movements (mean RT ~0.25s) compared to humans (~0.55s).
  • Drift Rate Reversal: Interestingly, humans had a higher drift rate ($k$) for motion, while monkeys had a higher drift rate for 3D. The authors attribute this to the monkeys' extensive, long-term training on the 3D task.

Limitations and Future Work

The primary limitation is the small sample size for the monkey subjects ($n=2$), which is common in primate research but makes it harder to generalize individual differences. Additionally, because the monkeys were much more highly trained than the humans, it's difficult to say if the "drift rate" differences are biological or simply a result of expertise.

The authors suggest that the next step is to use multi-site neurophysiological recordings to see if the distinct "integration windows" predicted by their models actually correspond to different patterns of neural firing in the parietal and frontal cortex.

Summary updated

Abstract

Abstract Decision-making models distil principles of information processing that underlie a range of cognitive functions. For the cases of motion detection in random-dot kinematograms and decisions about the rotation of 3D structure-from-motion cylinders, contributing neural processes have been localized to specific circuits in extrastriate area V5/MT on the basis of both causal and correlative evidence, suggesting a common decision path. Here, we arranged for humans and rhesus monkeys to make perceptual choices about these two stimulus types, indicating their decision with either hand or eye movements. In both species, the parameter distributions of a hierarchical Drift Diffusion Model (DDM) revealed systematic differences for stimulus type but not the different modes of response. States in a hidden Markov model also indicated distinct decision strategies, again for the two stimulus types but not for response mode. Although physiological evidence points to area V5/MT as a vehicle for relevant perceptual signals for both stimulus types, computational modelling of the present results reveal distinct decision processes, therefore predicting that different neural processes underlie judgements about motion and 3D-depth, consistently for humans and monkeys.
BibTeX
@misc{rangotis2024distinct,
  author = {Revan Rangotis and Sylwia Nowakowska and Peter Dayan and Andrew Parker and Ehsan Kakaei and Akinlolu Akande and Kristine Krug},
  title = {{Distinct decision processes for 3D and motion stimuli in both humans and monkeys revealed by computational modelling}},
  year = {2024},
  doi = {10.1101/2024.11.25.625189},
  howpublished = {bioRxiv},
}
Figure