This repository contains the code and models developed by Team -4R² for the UCSD SMASH & NSF HDR Hackathon, where we won second place in the Neural Forecasting Track. The project focuses on predicting monkey motor neuron activity using neural time-series data.
The challenge was inspired by a current NSF HDR global challenge and tasked participants with forecasting neural activity. Specifically, given the previous 10 time steps of neural data, the goal was to predict the next 10 time steps. Accurate forecasting of neural signals has applications in neuroscience research, brain-computer interfaces, and motor prosthetics.
Over the hackathon weekend, we explored multiple deep learning approaches to this time-series forecasting problem, including:
- LSTM (Long Short-Term Memory networks)
- GRU (Gated Recurrent Units)
- Temporal Convolutional Networks (TCN)
After extensive experimentation and performance comparison, we observed:
- LSTM and GRU achieved similar forecasting accuracy.
- GRU was selected as our final model due to slightly better performance as hyperparameters were tuned further.
- TCN provided competitive results but did not surpass the recurrent architectures for this dataset.
- Implementation of multiple neural network architectures for time-series forecasting.
- End-to-end workflow: data preprocessing → model training → evaluation of predictions.
- Performance evaluation based on next-step forecasting accuracy.