Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Deep learning in a bilateral brain with hemispheric specialization

Chandramouli Rajagopalan, David Rawlinson, Elkhonon Goldberg, Gideon Kowadlo

TL;DR

This work investigates whether a bilateral neural architecture, with the left hemisphere specializing in local features and the right in global features, can improve performance on a hierarchical image classification task. The authors implement bilateral CNNs (using ResNet-9 or VGG-11A) and induce specialization via dual objectives on fine and coarse labels, comparing against several baselines and ensembles. Results show that Bilateral-specialised models generally outperform single-hemisphere baselines, though a conventional 2-model unilateral ensemble often achieves the highest accuracy; gradient-based and cosine-similarity analyses reveal that the heads learn a weighted, nonlinear integration of complementary features from both hemispheres. The findings support the viability of bilateralism as an inductive bias for AI and offer insights into how neuro-inspired specialization can be leveraged for improved generalization and task-specific feature fusion.

Abstract

The brains of all bilaterally symmetric animals on Earth are divided into left and right hemispheres. The anatomy and functionality of the hemispheres have a large degree of overlap, but there are asymmetries, and they specialise in possesses different attributes. Other authors have used computational models to mimic hemispheric asymmetries with a focus on reproducing human data on semantic and visual processing tasks. We took a different approach and aimed to understand how dual hemispheres in a bilateral architecture interact to perform well in a given task. We propose a bilateral artificial neural network that imitates lateralisation observed in nature: that the left hemisphere specialises in local features and the right in global features. We used different training objectives to achieve the desired specialisation and tested it on an image classification task with two different CNN backbones: ResNet and VGG. Our analysis found that the hemispheres represent complementary features that are exploited by a network head that implements a type of weighted attention. The bilateral architecture outperformed a range of baselines of similar representational capacity that do not exploit differential specialisation, with the exception of a conventional ensemble of unilateral networks trained on dual training objectives for local and global features. The results demonstrate the efficacy of bilateralism, contribute to the discussion of bilateralism in biological brains, and the principle may serve as an inductive bias for new AI systems.

Deep learning in a bilateral brain with hemispheric specialization

TL;DR

This work investigates whether a bilateral neural architecture, with the left hemisphere specializing in local features and the right in global features, can improve performance on a hierarchical image classification task. The authors implement bilateral CNNs (using ResNet-9 or VGG-11A) and induce specialization via dual objectives on fine and coarse labels, comparing against several baselines and ensembles. Results show that Bilateral-specialised models generally outperform single-hemisphere baselines, though a conventional 2-model unilateral ensemble often achieves the highest accuracy; gradient-based and cosine-similarity analyses reveal that the heads learn a weighted, nonlinear integration of complementary features from both hemispheres. The findings support the viability of bilateralism as an inductive bias for AI and offer insights into how neuro-inspired specialization can be leveraged for improved generalization and task-specific feature fusion.

Abstract

The brains of all bilaterally symmetric animals on Earth are divided into left and right hemispheres. The anatomy and functionality of the hemispheres have a large degree of overlap, but there are asymmetries, and they specialise in possesses different attributes. Other authors have used computational models to mimic hemispheric asymmetries with a focus on reproducing human data on semantic and visual processing tasks. We took a different approach and aimed to understand how dual hemispheres in a bilateral architecture interact to perform well in a given task. We propose a bilateral artificial neural network that imitates lateralisation observed in nature: that the left hemisphere specialises in local features and the right in global features. We used different training objectives to achieve the desired specialisation and tested it on an image classification task with two different CNN backbones: ResNet and VGG. Our analysis found that the hemispheres represent complementary features that are exploited by a network head that implements a type of weighted attention. The bilateral architecture outperformed a range of baselines of similar representational capacity that do not exploit differential specialisation, with the exception of a conventional ensemble of unilateral networks trained on dual training objectives for local and global features. The results demonstrate the efficacy of bilateralism, contribute to the discussion of bilateralism in biological brains, and the principle may serve as an inductive bias for new AI systems.
Paper Structure (25 sections, 9 figures, 2 tables)

This paper contains 25 sections, 9 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Related work
  3. Methods
  4. Dataset
  5. Models and specialisation approach
  6. Common elements -- backbones and heads
  7. Bilateral network with specialisation
  8. Unilateral network with specialisation
  9. Unilateral network without specialisation
  10. Bilateral network without specialisation
  11. Ensembles
  12. Training procedure
  13. Visualising network operation
  14. Gradient camera (Grad-Cam) visualisation
  15. Feature analysis using cosine similarity
  16. ...and 10 more sections

Figures (9)

  • Figure 1: Bilateral architecture and baselines. Feature extractors are either ResNet-9 or VGG-11. The classifier heads are single layer fully connected networks, with output dimensions the same size as the number of classes. For brevity, if not specified, the network is unspecialised.
  • Figure 2: Accuracy of ResNet-based models. For brevity, if not specified, the network is unspecialised.
  • Figure 3: Accuracy of VGG-11-based models. For brevity, if not specified, the network is unspecialised.
  • Figure 4: Grad-Cam for Scenario 1: The bilateral-specialised (abbreviated to bilateral-sp in the figure) network is correct, left and right are incorrect. The bilateral network appears to adjust the area of focus even when both the hemispheres' features are situated external to the bosc.
  • Figure 5: Grad-Cam for Scenario 2: The bilateral-specialised network and the right are correct, the left is incorrect. Possibly due to the fact that there are multiple labels with wheels, the left failed to differentiate the tank by its wheels alone, however the bilateral model appears to overcome this, using the right's features.
  • ...and 4 more figures