courses:wshop:topics:tematy2023wiosna [IIS Wiki]

This is an old revision of the document!

Student:
Namespace in the wiki: FIXME
The goal of the project: Exploration of existing semantic catalogues for the manuscript collections
Technology: Semantic Web (RDF, SPARQL, ontologies), scientific papers reading/mail writing
Description: A survey of the systems used for the storage and management of manuscripts should be prepared (use the links below as a starting point) and then compared by presenting the content/assumptions/purpose, the semantic model underpinning the project and the practical possibilities for inferring/filtering/displaying collection items.
There will probably be no way to get to the details of the catalogue directly from the website, so one will need to look for scientific papers describing the systems, and if this still does not provide an answer then the people responsible for these projects should be contacted.
Links:
- Kalliope [DE]: https://kalliope.staatsbibliothek-berlin.de/en/index.html
- Fibula [PL]: http://info.filg.uj.edu.pl/fibula/en
- Goethe- und Schiller-Archiv [DE]: https://ores.klassik-stiftung.de/ords/f?p=401:1::::::
- Editionenportal [DE; still beta]: http://editionenportal.de/
- Wittgenstein Ontology Explorer [NO]: http://wab.uib.no/sfb/
- Wittgenstein Source [NO]: http://www.wittgensteinsource.org/

Student:
Namespace in the wiki: FIXME
The goal of the project: Add image modality to the explanation mechanism based on LUX
Technology: Python, Explainable AI, DeepNeuralNetworks,
Description: Explainable ai aim in providing human-readable explanations to model decisions (see more: Interpretable Machine Learning). Local uncertain explanations is a mechanism that generates rules that explains model decisions. It now only works for tabular data, and the goal of the project is to move it to image modality. In particular following milestones are expected to be accomplished:
- Use Zenit-CRP or GradCamto obtain concepts from image that will serve as conditional parts in the rule (you can use the VGG dataset that is given in the tutorial)
- Extract this concepts from image and based on the concepts IDs train LUX to explain different classes
- Present in visual rule-based form the explanations that will possibly allow to spot some biases in the dataset (like error in predicting garbage truck based on stop sign that appeared in the background most often)

Student:
Namespace in the wiki: FIXME
The goal of the project: Add image modality to the explanation mechanism based on LUX
Technology: Python, Explainable AI, DeepNeurralNetworks, AnomalyDetection,
Description: The goal of the project is explain Deep neural classifier in a rule-based manner, similarly to previous project. We will be working with real data from scientific project that our team is working on (See XPM). Following milestones will be required to accomplish the project:
- Use autoencoder delivered by our team (later in the semester use the auto-encoder developed by other student-team) to detect anomalies (dataset that we will be working with: Metro de Porto
- Build classifier (based on DNN) to predict the anomalies tagged by the AE.
- Use Grad-CAM to explain the classifier (obtain segments that are meaningful to detect anomaly)
- Train LUX on the segments defined by GRAD-Cam heat-maps:

Student:
Namespace in the wiki: problogcounterfactuals
The goal of the project: Create Counterfactual explanations of DeepNeuralNetworks with usage of probabilistic declarative programming
Technology: DeepProbLog, Python
Description: The goal of the project is to create counterfactual explanations (i.e. What I need to change in the input of the instance to change the prediction of the classifier to desired output. For instance, if the system predicts I cannot get a loan, I want to know what should I do (how change my customer profile) to get a loan. There are several methods fro that, but they most often are “blond” searchers. Therefore you may get a suggestion to change your gender, or your date of birth. To prevent that, we want to modulate the action the counterfactual search can perform in order to find the best one. The goal is to use probabilistic programming, where we can define a meta-procedure of generating counterfactual and allow the system to induce the concrete procedure itself (similar to inductive programming).

Links:
- DeepProbLog
- Dataset: Lets start with simple one (like Titanic, MNIST, etc.)

Student:
Namespace in the wiki: FIXME
The goal of the project: Prepare a script that will build meta-learnign dataset out of OpenML logs
Technology: Python, OpenML API
Description: The main goal of the project is to create a script that will fetch all of the runs/pipelines and dataset from OpenML platform and create a dataset out of it. The challenge is to transform pipeline definitions which are code snippets into logical components of machine-learning pipeline (including deep neural networks). Such a dataset will serve as a learn-to-learn dataset for meta-learning solutions.

Student:
Namespace in the wiki: FIXME
The goal of the project: Contribute to the InXAI framework by extending it with image modality
Technology: Python
Description: The goal of the project is to extend the InXAI framework to include image modality. In particular the milestones required to accomplish the project will be as follows:
- Implement segmentation mechanism for images
- Implement permutation mechanism for segmented image (e.g. replacing segments with average value of pixels, etc.)
- Obtain LIME/SHAP importance for the segments of the image
- Calculate metrics with InXAI (this part is already implemented, but need to be tested with Segmenters and Perturbers for images)

Links:
- InXAI
- You can start with: some-not-entirely-working-preliminary-work

Student:
Namespace in the wiki: FIXME
The goal of the project: Implement Oblique version of decision tree generation algorithm
Technology: Python
Description: The goal of the project is to extend the PyUID3 algorithm to use oblique split criterion.