This is a school project completed for the Cognition and Computation course at the University of Padova.
In this project, I investigated the process of visual concept learning with DBNs using simulations and analyses on FashionMNIST dataset. Visual concept learning is the process of acquiring knowledge about the visual features associated with different categories of objects. To do this, I performed different tasks, specifically:
- Analysis of internal representations using hierarchical clustering (dendograms) and feature visualization methods (receptive fields).
- Testing the model's representations at different levels of the hierarchy using linear read-outs.
- Visualizing confusion matrices and psychometric curves to examine the model's performance under varying perturbations.
- Evaluating the model's robustness against adversarial attacks.
Finally, I discussed the details of the implementation, the motives behind my choices, and critically evaluated the obtained results.