25 April 2024
10:00 Master's Defense IC3 Auditorium
Adapting autoencoders to reduce dimensionality and preserve information in the context of human activity recognition
Darlinne Hubert Palo Soto
Advisor / Teacher
Edson Borin
Brief summary
This work studies whether different dimensionality reducers can automatically discover (without human intervention) a data representation that is good and compact enough to preserve most of the information in time series data obtained from motion sensors (accelerometer and gyroscope), all in the context of the Human Activity Recognition (HAR) task. Therefore, transformations in the nature of the data, such as the Fourier transform or similar, are not allowed. The reducers studied were two autoencoder variants, and UMAP. Adding topological features as a regularization term is a new strategy, while adding convolutions is a common strategy in several works. On the other hand, UMAP showed good results when it was previously tested with data representations in the frequency domain. A data representation with a small loss of information should allow simple classification models (such as Random Forest, Support Vector Machines, and K-Nearest Neighbors) to achieve similar or better results than using the original data representation. Consequently, these improvements should also show the ability or otherwise of the reducer to produce good data representations. To ensure maximum performance of all reducers (and thus fairness in the comparison process), a hyperparameter search was performed for each and then they were contrasted with unseen data. The results confirm that generic autoencoders have trouble discovering good representations, while convolutional autoencoders are better suited to the task. Furthermore, adding topology to autoencoders generally resulted in lower accuracies. But even this topological variant was superior to UMAP. On the other hand, looking at dimensionalities when searching for hyperparameters revealed that improvements in accuracy were not necessarily linked to high dimensionalities. Instead, a small percentage of the original dimensions was sufficient to achieve the reducers' maximum performance.
Examination Board
Edson Borin IC / UNICAMP
Flavia Cristina Bernardini IC / UFF
Paula Dornhofer Paro Costa FEEC / UNICAMP
Hélio Pedrini IC / UNICAMP
Daniel Carlos Guimarães Pedronette IGCE / UNESP
Pedro Mário Cruz e Silva NVIDIA