A Partially Shared Latent Neural Space for Deductive Reasoning

Read more

• Paper: link coming soon
• Code and data: GitHub link coming soon
• Project page: link coming soon

If you are interested in reasoning, cognitive neuroscience, latent-variable modeling, or task-based functional connectivity, I would be happy to discuss.

Introduction

Most fMRI studies of reasoning ask a familiar question:

Which brain regions become more active when people reason?

That question matters. But it may not be the whole story.

Reasoning is unlikely to live in a single isolated region. When someone solves a deductive problem, many brain areas respond together: frontal regions, parietal regions, subcortical systems, and networks involved in control, memory, and representation.

So in this project, we ask a different question:

Can reasoning-related brain activity be explained by a small number of shared latent neural factors?

Instead of asking only which regions activate, we ask whether trial-wise brain activity can be summarized by lower-dimensional neural factors.

The problem: “where is reasoning?” may be too narrow

Classic task-fMRI studies usually compare conditions:

• reasoning vs. baseline
• one reasoning task vs. another
• valid vs. invalid arguments
• easy vs. difficult trials

These contrasts are useful. They show which regions are more active in one condition than another.

But a reasoning trial is not just a peak in one region.

It is a distributed activity pattern across many brain areas. If these areas covary across trials, then the neural response to reasoning may be more compact than it first appears.

High-dimensional brain activity may contain low-dimensional structure.

What we study

We focus on two forms of deductive reasoning:

Syllogistic reasoning involves arguments such as:

All A are B. All B are C. Therefore, all A are C.

Transitive reasoning involves relational chains such as:

A is greater than B. B is greater than C. Therefore, A is greater than C.

Both require inference. But they may rely on partly different representational demands.

That makes them a useful test case: are there neural factors shared across reasoning tasks, and are some factors task-specific?

The key idea

For each trial, we estimate an ROI-level activity pattern:

one trial → one vector of brain-region responses

This gives us a trial-by-region matrix.

Then we use latent-factor methods such as PCA and factor analysis to ask whether many regional responses can be summarized by a smaller number of neural components.

Trial-wise ROI activity is modeled as a matrix. Latent-factor methods ask whether this matrix contains stable lower-dimensional structure.

The goal is not to immediately call a factor “logic,” “working memory,” or “relational integration.”

Instead, we ask:

• Is the factor stable?
• Does it generalize across tasks?
• Does it distinguish task conditions?
• Does it relate to behavior?
• Does it map onto meaningful brain networks?

Why functional connectivity matters

Activation tells us which regions respond.

Functional connectivity asks whether regions coordinate with each other.

As a complementary analysis, we test whether latent activation factors relate to task-based network coordination. In other words:

Do low-dimensional reasoning factors reflect coordinated activity among brain networks?

The project links trial-wise activation factors with functional connectivity and behavioral performance.

Why this matters

This project moves beyond the question:

Where is reasoning activated?

and asks:

What low-dimensional neural structure supports reasoning across tasks?

The broader goal is to connect deductive reasoning, trial-wise fMRI, latent-variable modeling, and network neuroscience.

Reasoning may not be best understood as a list of active regions.

It may be better understood as a structured neural state space.