AdjointBackMap: Reconstructing effective decision hypersurfaces from CNN layers using adjoint operators.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Understanding the internal decision-making processes of Convolutional Neural Networks (CNNs) is crucial for their reliable deployment.
• Inverting the complex functions learned by CNNs is generally an ill-posed problem, hindering interpretability.

Purpose of the Study:

• To develop a method for reconstructing the effective input space representation of individual units within a CNN.
• To visualize and understand the decision boundaries and input sensitivities of CNN units.

Main Methods:

• The study proposes a novel method utilizing adjoint operators to reconstruct a hypersurface for each CNN unit (excluding the first convolutional layer).
• This reconstructed hypersurface represents the unit's effective influence in the input space and is visualized directly.

Main Results:

• The reconstructed hypersurfaces accurately predict the output of individual CNN units when applied to input images.
• The visualization demonstrates that CNN unit decisions are highly input-dependent.
• Reconstructed hypersurfaces exhibit significant sensitivity to adversarial noise, correlating with known CNN vulnerabilities.

Conclusions:

• The adjoint operator method provides a powerful tool for interpreting CNN unit behavior and visualizing their function in the input space.
• The findings offer new insights into the susceptibility of CNNs to adversarial attacks by highlighting the sensitivity of their internal representations.
• This approach contributes to the broader goal of developing more robust and interpretable deep learning models.