Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Cognitive Neural Architecture Search Reveals Hierarchical Entailment

Lukas Kuhn, Sari Saba-Sadiya, Gemma Roig

TL;DR

This work investigates whether hierarchical entailment in the primate ventral stream can emerge when CNN architectures are evolved to align with late-ventral representations. It employs an evolutionary neural architecture search (NAS) framework that optimizes architecture for brain-alignment using fMRI data from the NSD dataset, evaluating fitness via ridge-encoded fMRI predictions. The evolved architectures (notably EvoV2, EvoV4, EvoIT) achieve brain-alignment superior to several manually designed CNNs and reveal subnetworks indicative of a hierarchical processing cascade, with IT-aligned models containing components predictive of earlier regions. Training the best-evolved architectures on CIFAR-10 shows competitive classification performance and highlights that gradient-based learning can modulate brain-alignment metrics, suggesting NAS as a powerful tool for computational cognitive neuroscience with potential to reduce manually designed architectures.

Abstract

Recent research has suggested that the brain is more shallow than previously thought, challenging the traditionally assumed hierarchical structure of the ventral visual pathway. Here, we demonstrate that optimizing convolutional network architectures for brain-alignment via evolutionary neural architecture search results in models with clear representational hierarchies. Despite having random weights, the identified models achieve brain-alignment scores surpassing even those of pretrained classification models - as measured by both regression and representational similarity analysis. Furthermore, through traditional supervised training, architectures optimized for alignment with late ventral regions become competitive classification models. These findings suggest that hierarchical structure is a fundamental mechanism of primate visual processing. Finally, this work demonstrates the potential of neural architecture search as a framework for computational cognitive neuroscience research that could reduce the field's reliance on manually designed convolutional networks.

Cognitive Neural Architecture Search Reveals Hierarchical Entailment

TL;DR

This work investigates whether hierarchical entailment in the primate ventral stream can emerge when CNN architectures are evolved to align with late-ventral representations. It employs an evolutionary neural architecture search (NAS) framework that optimizes architecture for brain-alignment using fMRI data from the NSD dataset, evaluating fitness via ridge-encoded fMRI predictions. The evolved architectures (notably EvoV2, EvoV4, EvoIT) achieve brain-alignment superior to several manually designed CNNs and reveal subnetworks indicative of a hierarchical processing cascade, with IT-aligned models containing components predictive of earlier regions. Training the best-evolved architectures on CIFAR-10 shows competitive classification performance and highlights that gradient-based learning can modulate brain-alignment metrics, suggesting NAS as a powerful tool for computational cognitive neuroscience with potential to reduce manually designed architectures.

Abstract

Recent research has suggested that the brain is more shallow than previously thought, challenging the traditionally assumed hierarchical structure of the ventral visual pathway. Here, we demonstrate that optimizing convolutional network architectures for brain-alignment via evolutionary neural architecture search results in models with clear representational hierarchies. Despite having random weights, the identified models achieve brain-alignment scores surpassing even those of pretrained classification models - as measured by both regression and representational similarity analysis. Furthermore, through traditional supervised training, architectures optimized for alignment with late ventral regions become competitive classification models. These findings suggest that hierarchical structure is a fundamental mechanism of primate visual processing. Finally, this work demonstrates the potential of neural architecture search as a framework for computational cognitive neuroscience research that could reduce the field's reliance on manually designed convolutional networks.

Paper Structure

This paper contains 12 sections, 2 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Method
  3. Search Space
  4. Fitness Evaluation
  5. Genetic Operators
  6. Dataset
  7. Baseline Models
  8. Experiments and Results
  9. Evolutionary Search for Cognitive Models
  10. Representational Hierarchy in Evolutionary Cognitive Models
  11. Training Evolutionary Cognitive Models for Image Classification
  12. Discussion and Future Work

Figures (2)

  • Figure 1: The evolutionary neural architecture search framework: Starting with a generation of neural architectures, for each model embeddings of all images in the shared NSD dataset are extracted. A ridge regression is then trained to predict the recorded fMRI activity and the correlation coefficient between the predicted and ground truth fMRI is calculated. The models are evaluated based on the mean correlation for all subjects, and the bottom 50% are eliminated. Finally genetic operations are used to repopulate the models for the next generation.
  • Figure 2: Left: The average rewards for the architectures optimized to predict V2, V4, and IT brain activations in each generation. Right: The optimal EvoIT architecture.