This stage of the project is devoted to refining a rough exploratory study into a publishable paper. Based on some reccomendations, the code has been optimized for quantification. I managed to build a full data pipeline indexing the whole CAT2000 dataset (N=2,000), consisting of the training and testing subsets and, thus, producing an unbiased evaluation of the models.

Technical achievements that i have achieved are based on building an analysis/ folder, which will be used as the evidence storage. It includes automated scripts measuring traditional saliency metrics, namely Pearson Correlation Coefficient (CC) and Similarity (SIM) scores, together with the tools of visualization of extremal cases of the experiment’s performance. Thus, we have constructed a transparent method of transparency analysis.

We seek to offer a geometric rationale for the successful application of the concept of transfer learning. The main idea is to use activation maps as points of Riemannian manifolds and calculate their features, such as smoothness of the manifold, and show that pre-training results in semantically aligned features space. Thus, i want to establish the connection between geometric properties of the activation manifold and predictive metrics, making a significant contribution to saliency research.

demo link: https://github.com/jedilikestocode/NHSJS-Research-Code