Skip to content

MrRezaeiUofT/Alpha-Alternator

 
 

Repository files navigation

Alpha Alternator: Dynamic Adaptation To Varying Noise Levels In Sequences Using The Vendi Score For Improved Robustness and Performance

arXiv

The Alpha Alternator is a novel generative model designed for time-dependent data, dynamically adapting to varying noise levels in sequences. Unlike state-of-the-art dynamical models such as Mamba, which assume uniform noise across sequences, the Alpha Alternator leverages the Vendi Score (VS) to adaptively adjust the influence of sequence elements on predicted future dynamics at each time step.

Alpha Alternator Noise Robustness

Key Features

  • Adaptive Noise Handling: Dynamically balances reliance on input sequences and latent history based on learned parameters.
  • Vendi Score (VS) Integration: Uses a similarity-based diversity metric to assess sequence informativeness.
  • Alternator Loss Optimization: Enhances robustness through observation masking and targeted loss minimization.
  • Superior Performance: Outperforms state-space models and Alternators in trajectory prediction, imputation, and forecasting.

Methodology

The Alpha Alternator dynamically adjusts its prediction strategy based on a learned parameter.

Training involves observation masking to simulate diverse noise levels and Alternator loss minimization to improve model resilience.

Installation

To set up the environment, install the required dependencies:

pip install -r requirements.txt

Monash and Neural Datasets

The datasets used in this research can be downloaded from:

These datasets are available in both MATLAB and Python formats.

Results

Experimental results demonstrate that the Alpha Alternator achieves state-of-the-art performance in neural decoding and time-series forecasting, surpassing existing Alternators and state-space models.

Citation

If you use the Alpha Alternator model in your research, please cite:

@article{rezaei2025alpha,
  title={The Alpha Alternator: Dynamic Adaptation to Varying Noise Levels in Sequences Using the Vendi Score for Improved Robustness and Performance},
  author={Rezaei, Mohammad R. and Dieng, Adji Bousso},
  journal={arXiv preprint arXiv:2502.04593},
  year={2025}
}

License

This project is licensed under the MIT License. See the LICENSE file for details.

About

The Alpha Alternator is a novel generative model designed for time-dependent data, dynamically adapting to varying noise levels in sequences.

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

 
 
 

Contributors

Languages

  • Python 100.0%