This project investigates a modified version of PAR-Net, a dual-stream architecture that combines a Convolutional Neural Network with an Echo State Network reservoir and temporal attention.
Unlike the original architecture, the implementation preserves the full reservoir-state trajectory and aligns it with the CNN feature sequence before applying multi-head attention. Selected reservoir parameters are also made trainable, allowing the model to adapt its temporal dynamics to the classification task.
The architecture was evaluated on the FordA and FordB time-series classification datasets and compared against:
- A baseline CNN
- A CNN with multi-head attention
- The modified CNN–ESN–attention architecture
Experiments used the official dataset splits and five-fold cross-validation. Results showed that the reservoir-enhanced model achieved modest, consistent improvements under clean conditions, while the attention-only model was generally more stable under noisy distribution shifts.
211591_CONANAN_CSCI214_FinalPaper.pdf— project paper211591_CONANAN_CSCI214_FinalPaper_Code.ipynb— model implementation and experiments
Python, PyTorch, NumPy, scikit-learn, pandas, and matplotlib
Raul Jarod Conanan
Developed as the final project for CSCI 214.