What it is:

Deep Learning-based Representational Similarity Analysis for decoding numerosity representations from EEG.

What we did:

  • Trained compact CNN (EEGNeX) decoders on every numerosity pair (1-6) to measure brain representational distances.
  • Generated Representational Dissimilarity Matrices (RDMs) that reveal the PI-to-ANS transition and divisibility effects.
  • Controlled for visual confounds (pixel area) via RSA-style partial correlations and deterministic LOSO splits.
  • Nested, subject-aware CV (outer LOSO, inner GroupKFold) so Optuna, early stopping, and refits stay leak-free. outer folds ensemble.
  • View project on GitHub
Read Full Project Details...

EEG Deep Learning for Numerical Cognition

A deep learning pipeline for decoding numerical representations from 128-channel EEG data. This project implements Deep Learning-based Representational Similarity Analysis (RSA) to map the neural geometry of the Parallel Individuation (PI) and Approximate Number System (ANS).

Scientific Motivation

The Two-Systems Hypothesis

Cognitive neuroscience proposes two distinct systems for processing quantities:

  1. Parallel Individuation (PI): Processes small numbers via rapid, precise “object files.”
  2. Approximate Number System (ANS): Processes larger numbers via magnitude estimation, where precision follows Weber’s law.

This project asks

Can we use Convolutional Neural Networks (CNNs) to map the representational state space of numerosity?

Specifically:

  • The Boundary: Can we reveal patterns that suggest a clean divide between ‘small’ and ‘large’ numbers?
  • The Structure: Is there a unique representational geometry within the small-number range?
  • Grouping: Does the brain use grouping mechanisms to represent composite numbers?

Study Background

This project analyzes data from a numerical oddball task (N=24 adults, 6,480 trials). Participants viewed dot arrays while EEG was recorded. We apply Deep Learning RSA to decode fine-grained spatiotemporal patterns from raw, single-trial data.

What This Pipeline Does

This repository implements a pipeline that uses a compact CNN (EEGNeX) as a distance metric for Representational Similarity Analysis.

Representational Similarity Analysis (RSA):

  1. Train neural networks to distinguish every possible pair of numerosities.
  2. Use decoding accuracy as the measure of representational dissimilarity.
  3. Construct a Representational Dissimilarity Matrix (RDM) to visualize the neural geometry.
  4. Control for visual confounds (pixel area) using partial correlation analysis.

Findings

Uncovering the Neural State Space

By projecting our Deep Learning RDM into 2D space (Multidimensional Scaling), we uncovered a non-linear architecture of number processing in the adult brain.

1. System Distinctness & Boundary:
…

2. Structure Within Object Tracking:
The subitizing range is not composed of uniformly distinct slots. Numerosities 2-4 form a similarity cluster.

3. The Divisibility Effect:
High confusion between numerosities 5 and 6.

Robustness to Visual Confounds

We performed a control analysis to ensure these results were not driven by low-level visual features.

Supported Workflows

The pipeline supports end-to-end RSA execution:

  • sa_binary: Trains pairwise classifiers across 10 random seeds using Leave-One-Subject-Out (LOSO) cross-validation.
  • sa_pixel_control: Performs post-hoc statistical analysis to regress out pixel confounds from the brain RDM.
  • generate_rsa_tables: Produces publication-ready LaTeX tables of pairwise decoding statistics (Holm-corrected).

Features

  • Leak-Free Validation: Subject-aware splits ensure no participant data appears in both train and test.
  • Constitutional Rigor: All parameters must be explicitly specified via YAML.
  • Automated RSA: End-to-end scripts for training, RDM generation, and Multidimensional Scaling (MDS) visualization.
  • Explainable AI: Integrated Gradients highlight spatiotemporal feature importance.
  • Statistical Rigor: Deterministic seeding, permutation testing, and partial correlation analysis for confounds.

  • Full Provenance: Every run logs model class, library versions, hardware, and seeds.

    </div>

</details>