| Meno: | Hana |
|---|---|
| Priezvisko: | Hladíková |
| Názov: | Analysis of sparse echo-state network |
| Vedúci: | prof. Ing. Igor Farkaą, Dr. |
| Rok: | 2026 |
| Kµúčové slová: | reservoir computing, Echo State Networks, SpaRCe, adaptive thresholding, time series classification |
| Abstrakt: | This master's thesis focuses on the analysis, reimplementation, and experimental evaluation of the SpaRCe model, which represents a sparse extension of the Echo State Network architecture based on adaptive threshold representations. The thesis presents the theoretical foundations of reservoir computing, sparsity mechanisms in neural networks, and non-dissipative reservoir architectures designed for processing long-term temporal dependencies. As part of this work, the SpaRCe model was completely reimplemented in the PyTorch framework. The implementation was validated on the standardized MNIST and pMNIST datasets, achieving results comparable to those of the original implementation. Additional experiments analyzed the relationship between the initial sparsity parameter, the sparsity of the learned representations, and the classification performance of the model. The experimental part of the thesis included the evaluation of the SpaRCe model on four standardized benchmark datasets commonly used in non-dissipative reservoir computing research. The obtained results confirmed the competitive classification performance of the model and further showed that the influence of the initial sparsity parameter on classification performance depends on the characteristics of the processed dataset. To support these experiments, a reproducible experimental framework was developed, providing automated data preprocessing, hyperparameter configuration generation, repeated evaluation with multiple random initializations, result aggregation, and execution on HPC infrastructure using the SLURM workload manager. The developed framework provides a reliable and extensible foundation for future research on sparse and non-dissipative reservoir computing architectures. |
Súbory diplomovej práce:
| kody.zip |
| praca.pdf |
Súbory prezentácie na obhajobe: