Research Article Open Access

Visual Explainability through Layer Conductance Using the Gradient Ascent Method

Abstract

The significant strides of deep neural network architectures have driven impressive performance gains, but has also introduced greater complexity and opacity within hidden layers, presenting significant challenges for interpreting deep learning models. Various visual explainability techniques have emerged, such as Grad-CAM, Layer-wise Relevance Propagation (LRP), Saliency Maps, and DeepLIFT have been widely used to highlight important regions in an image, attributing these regions to the model's prediction. In this article, we propose a novel approach based on Layer conductance and gradient ascent, typically used to quantify neuron contributions in hidden layers, can also serve as a pathway to visual interpretability. Layer conductance is utilized here in a novel context: Rather than focusing on analyzing hidden units, we extend this approach to examine the impact at the input pixel level, as another gradient-based approach.

Amine Baazzouz*, Jaouad Dabounou

To read the full article Download Full Article | Visit Full Article