Name: MATHEUS BECALI ROCHA
Publication date: 19/06/2024
Examining board:
Name |
Role |
|---|---|
| GUILHERME PALERMO COELHO | Examinador Externo |
| RENATO ANTONIO KROHLING | Presidente |
| VINICIUS FERNANDES SOARES MOTA | Examinador Interno |
Summary: Variational Autoencoders (VAEs) are generative models known for learning compact and continuous latent representations of data. Despite their effectiveness in various applications, using these representations for classification tasks still presents a challenge. Traditionally, this approach involves concatenating the mean and variance vectors and feeding them into a neural network. In this dissertation, a new approach for Variational Autoencoders, named VAE-GNA, was developed, which incorporates Gaussian neurons in the latent space along with attention mechanisms. These neurons directly process the mean and variance values through a modified sigmoid function, optimizing the VAE's training in feature extraction in synergy with the classification network. Additionally, the application of both additive and multiplicative attention mechanisms was proposed to enhance the model's capabilities. To evaluate the methodology, three experiments were conducted: the first employs Raman urine data for bladder cancer detection; the second uses near-infrared (NIR) data, also from urine, for the detection of various types of cancer; and the third uses NIR data for skin cancer detection. The results obtained show the effectiveness and applicability of the approach for spectral data classification.
