<aside> 👉 The Self-Explainable Deep Neural Networks project aims to make AIs able to explain themselves without the need for external tools. We aim to achieve the goal by (1) developing novel deep neural networks that are easier to understand than the current models, (2) developing layers that can be added to current networks to improve their transparency, or (3) generalizing current self-explainable deep neural networks on novel domains and tasks.

</aside>

Spring 2025 Auditor Project

The Spring 2025 Auditor Project for the self-explainable deep neural networks project will focus on analyzing and correcting the behavior of self-explainable Deep Neural Networks based on memory. As part of the project, you will generate the explanation images for the wrong predictions and propose a "strategy” (or alternative memory samples) to correct these errors.

Week 1: AIEA Lab Onboarding

Week 2: Project Onboarding (Get the reference code repository running locally and generate and visualize the explanation figures (memory images) for the pre-trained model on the test dataset )

Week 3: Nautilus

Week 4: Get the reference code repository running on Nautilus in a headless environment (you will need to be well acquainted with the terminal). Document the process.

• Note that this task is mandatory if your workstation cannot run the “standard” configuration, as completing the next steps will not be possible otherwise.

Week 5: Read the paper and write a small report summarizing it in your own words.

Week 6: Modify the script to compute and print the indices of all the wrong predictions (errors). Choose 20 indices among them and communicate the selection to the teaching team.

Week 7: Modify the script to generate an explanation figure only for the 20 selected indices at the previous step and enrich the figures with additional information (e.g., true class).

Week 8: For each of the 20 selected wrong predictions, test different memory sets as input alongside them (at least 3 memory sets per prediction). What do you observe?

Week 9: Try to find, for as many predictions as possible (min 2), a memory set such that when fed to the network alongside the image, it allows the model to return the correct prediction (and thus it “corrects” the model) for that image. Comment on your strategy for selecting the memory set.

Week 10: Auditor Offboarding + final report writing (5 pages max)

Policy on Bonus Point Replacements: You can use bonus points to replace a maximum of two deliverables of this project. You may use bonus points to replace Week 4, Week 5, or only one between Week 8 and Week 9

• For example, missing deliverables 4 and 5 and collecting at least 2 bonus points to compensate for them is fine. Conversely, if you miss both the 8 and 9 deliverables, you cannot complete the course even if you have more than 2 bonus points.