Biological convolutions improve DNN robustness to noise and generalisation

Area of Science:

• Computer Vision
• Computational Neuroscience
• Artificial Intelligence

Background:

• Deep Convolutional Neural Networks (DNNs) show high accuracy in image classification.
• DNNs are proposed as models for the primate visual system due to architectural similarities.
• DNNs may rely on 'shortcut learning' (e.g., texture over shape), limiting generalization and robustness.

Purpose of the Study:

• To investigate if incorporating biological filter banks can mitigate shortcut learning in DNNs.
• To enhance DNNs' internal representations and tolerance to noise and out-of-distribution images.
• To improve DNNs' generalization capabilities for real-world vision tasks.

Main Methods:

• Implemented fixed biological filter banks, specifically Gabor filters, within DNN architectures.
• Trained and tested DNNs with and without these filter banks on various image datasets.
• Evaluated network performance on standard benchmarks and novel out-of-distribution image sets.

Main Results:

• Adding Gabor filter banks constrained DNNs, reducing reliance on shortcut learning.
• Networks with filter banks developed more structured internal representations.
• DNNs with filter banks showed increased tolerance to noise and improved generalization (20-35% accuracy gain on novel datasets).

Conclusions:

• Biological filter banks can guide DNNs to learn more robust and generalizable representations.
• Incorporating primate visual system properties, like Gabor filtering, enhances DNN performance in challenging conditions.
• This approach offers a pathway for developing AI systems that better emulate human visual perception and real-world adaptability.