Generative and Deep Learning

In the following you can find a list of final project topics: they all require

1. implementing a method/application from literature
2. validating it as appropriate
3. reporting the model details, implementation and validation in a written report
4. summarizing the content of the report in a 15 minutes presentation to be delivered on the oral day

The report should be about 10 pages in length, single column and delivered in PDF.

The presenentation should be not more than 15 slides, also delivered in PDF.

All projects are meant to be developed individually. The use of coding agents is allowed, but should be described clearly in the report and in the presentation.

Students willing to pursue a final project should contact the instructor to agree on the exact scope of the project and the data to be used.

Final projects list

1. Confront a constraint-based, search-and-score and hybrid BN structure learning method, using at least 5 different datasets from the BnRep repository
2. Implement a Graph Topic Model and train/validate it on two graph datasets of your choice (from two different domains) 
3. Implement a simple mobile application for recognizing and transcribing text from the camera stream: to this end you will need to implement and train an efficient neural network (of your choice) for image data. You can consider using the letter recognition dataset to train it.
4. Implement a VAE for MIDI Music data, training it on the Lakh MIDI Dataset, and showing both generation capabilities as well as latent space manipulations.
5. Implement a flow-matching generative models for images, trained and validated on MNIST and CIFAR 100 datasets.
6. Implement the identifiable VAE (Khemakhem et al, 2020) and replicate results on synthetic data. Then, subsample the 3DShapes dataset to introduce correlations among shape and a subset of the remaining features, i.e., P(z|u='cube')≠P(z|u='sphere'). Using the shape as an auxiliary variable, investigate the behavior of iVAE and compare it to VAE/β-VAE.
7. Implement Ada-GVAE (Locatello et al, 2020) and replicate results on at least two datasets from the paper. Then, study the problem of training the model when only a portion of the dataset contains non i.i.d. pairs of observations. Report results on disentanglement for increasing portion of paired samples.
8. Implement the Anti-Symmetric Deep Graph Network approach to make model XXX propagate information long range: validate on the ECHO-synth part of the ECHO Benchmark.
9. Implement two different generative models for graph data of your choice (e.g. autoregressive, diffusion-based, VAE, ...) and validate them on QM9 chemical graphs
10. Implement a neural-based reinforcement learning agent of your choice (e.g. DQN) and experiment training it on at least two simulation environments available (e.g. Atari, Gymnasium, more..)