14May2026
09:30 Master's Defense room 85 of IC2
Topic on
Comparative Study of Autoencoders: Reconstruction, Generation, and Robustness of the Latent Space
Student
Wilson Bagni Júnior
Advisor / Teacher
Zanoni Dias - Co-advisor: Hélio Pedrini
Brief summary
This work presents a comprehensive comparative study between five autoencoder architectures: Fully Connected Autoencoder (FCAE), Convolutional Autoencoder (CAE), Variational Autoencoder (VAE), Adversarial Autoencoder (AAE), and RealNVP Autoencoder (RealNVP-AE). The central objective was to investigate the existing trade-offs between reconstruction fidelity, latent space organization, and generative capacity of these models. The methodology adopted a rigorous experimental approach using the Street View House Numbers (SVHN) database as the main scenario and the Synthetic Digits (SD) database for cross-validation, varying the dimensionality of the latent space between 16 and 1.536 dimensions. The results showed that deterministic models (CAE and FCAE) outperformed stochastic models in reconstruction quality and adaptation to new domains, while the probabilistic VAE model demonstrated superiority in the statistical validation of data generation, establishing itself as the most balanced architecture for generative tasks, despite producing reconstructions with less sharpness. In contrast, the adversarial approach (AAE), despite its robust theoretical foundation, presented generation metrics lower than expected for models that seek to regularize the latent space, thus highlighting the practical difficulty of training this model. One of the main findings was the structural collapse of RealNVP-AE in high dimensions, indicating scalability limitations for this class of flow-based models. Additionally, it was observed that robustness to noise in the latent space is more associated with the preservation of data topology than with the imposition of distributions. It is also concluded that increasing dimensionality does not always imply performance gains, suggesting that intermediate configurations offer the best balance between computational efficiency and representational performance.
Examination Board
Headlines:
| Zanoni Dias | IC / UNICAMP |
| Alexandre Mello Ferreira | EEP |
| Marcelo da Silva Reis | IC / UNICAMP |
Substitutes:
| Rafael de Oliveira Werneck | IC / UNICAMP |
| Ana Estela Antunes da Silva | FT / UNICAMP |