Publicação
SERS Spectra Classification using Deep Learning
| Resumo: | Recently, wines became a carefully engineered mixture of components. The particular- ities of each mixture are what make wines unique and, in order to make this analysis of mixtures and combinations, techniques of thorough identification of components are required. Such techniques, like SERS, produce very detailed information, however, that information can be hard to interpret. Given the large number of possible samples and the endless variations in the wines’ compositions, the interpretation becomes more com- plicated. The main goal of this study is to find relations within the characteristics and com- positions of a variety of Portuguese white wines. We propose the use of Deep Learning for the task of feature extraction and dimensionality reduction of the SERS spectra of the Portuguese white wines, in order to find a representation that is suitable for analysis by simple predictive algorithms. Deep Neural Networks have been proven to be very useful in finding simpler representations of complex data, while improving the relevance of the extracted features and preserving their structure, whilst simultaneously giving an insight on the characteristics the network considered most relevant for that represen- tation. Autoencoders are Deep Neural Networks especially designed for learning these representations from learned patterns in the data, with the sole purpose of making the representations sufficiently informative to be reconstructed back to its original state. In the presented thesis, we attempted to use traditional Autoencoders (AE) for feature extraction and dimensionality reduction prior to classification and clustering. With the autoencoders, it was not possible to generate representations that showed a clear distinc- tion between wine regions, types or grape variety, but were able to successfully generate representations that demonstrated the distinction between the different wines, meaning that this technique can be useful to provide understanding on the characteristics that distinguish each wine. |
|---|---|
| Autores principais: | Lucas, Matilde Alexandra Damas |
| Assunto: | Deep Learning Deep Neural Networks Autoencoder Dimensionality Reduction Feature Extraction urface Enhanced Raman Spectroscopy (SERS) |
| Ano: | 2023 |
| País: | Portugal |
| Tipo de documento: | dissertação de mestrado |
| Tipo de acesso: | acesso aberto |
| Instituição associada: | Universidade Nova de Lisboa |
| Idioma: | inglês |
| Origem: | Repositório Institucional da UNL |
| Resumo: | Recently, wines became a carefully engineered mixture of components. The particular- ities of each mixture are what make wines unique and, in order to make this analysis of mixtures and combinations, techniques of thorough identification of components are required. Such techniques, like SERS, produce very detailed information, however, that information can be hard to interpret. Given the large number of possible samples and the endless variations in the wines’ compositions, the interpretation becomes more com- plicated. The main goal of this study is to find relations within the characteristics and com- positions of a variety of Portuguese white wines. We propose the use of Deep Learning for the task of feature extraction and dimensionality reduction of the SERS spectra of the Portuguese white wines, in order to find a representation that is suitable for analysis by simple predictive algorithms. Deep Neural Networks have been proven to be very useful in finding simpler representations of complex data, while improving the relevance of the extracted features and preserving their structure, whilst simultaneously giving an insight on the characteristics the network considered most relevant for that represen- tation. Autoencoders are Deep Neural Networks especially designed for learning these representations from learned patterns in the data, with the sole purpose of making the representations sufficiently informative to be reconstructed back to its original state. In the presented thesis, we attempted to use traditional Autoencoders (AE) for feature extraction and dimensionality reduction prior to classification and clustering. With the autoencoders, it was not possible to generate representations that showed a clear distinc- tion between wine regions, types or grape variety, but were able to successfully generate representations that demonstrated the distinction between the different wines, meaning that this technique can be useful to provide understanding on the characteristics that distinguish each wine. |
|---|