On the role of feedback in image recognition under noise and adversarial attacks: A predictive coding perspective

Area of Science:

• Neuroscience and Machine Learning
• Computational Vision

Background:

• Brain-inspired machine learning is increasingly explored for computer vision tasks.
• The functional role of top-down feedback connections in convolutional networks remains unclear.

Purpose of the Study:

• To investigate the functional benefits of top-down feedback connections in convolutional neural networks (CNNs) for object recognition under noisy conditions.
• To determine how and when predictive coding dynamics improve network performance.

Main Methods:

• Implemented Predictive Coding (PC) dynamics via feedback connections in deep convolutional networks (CNNs).
• Trained networks for reconstruction or classification of clean images.
• Optimized and interpreted hyperparameters controlling recurrent dynamics to assess the role of predictive feedback.

Main Results:

• Networks implementing PC dynamics showed significantly higher accuracy than equivalent forward networks.
• Networks increasingly relied on top-down predictions as noise levels increased.
• In deeper networks, the reliance on top-down predictions was most prominent in lower layers.

Conclusions:

• Predictive feedback connections computationally enhance object recognition accuracy, particularly in noisy environments.
• Top-down predictions improve the robustness of vision models, offering insights for both neuroscience and machine learning.
• The study confirms the computational role of feedback connections in sensory systems.