Projects/BCI-FDA: Brain-computer interface with Functional data analysis

Brain-computer interface with Functional data analysis Human behavioral analysis and forecasting requires models that have to predict target variables of complex structures. We develop PLS and CCA (Projection to latent space and Canonic correlation analysis) methods towards the Multiview with continuous-time data representation.

Generative time series decoding models

1. The goal is to create a generative state-space model for BCI
2. The impact is to boost the behavioral classification quality by decision-rejecting
3. The principle if a generated pattern does not belong to one of the expected patterns (one-class classification) we reject the decision
4. The plan
   1. create the simplest generative model for selected data
   2. apply SSM (state-space model) principles to make CCA (canonic-correlation analysis)
   3. introduce classification model and decision-rejecting criterion
   4. compare quality

References

1. Direct Discriminative Decoder Models for Analysis of High-Dimensional Dynamical Neural Data by M.R. Rezaei et al. 2022 [DOI https://doi.org/10.1162/neco_a_01491]
2. Deep Direct Discriminative Decoders for High-dimensional Time-series Data Analysis by M.R. Rezaei 2023 (NTDB)
3. Decoding Hidden Cognitive States From Behavior and Physiology Using a Bayesian Approach by Ali Yousefi et al. DOI
4. Bayesian Decoder Models with a Discriminative Observation Process by M.R. Rezaei et al. 2020 DOI
5. Deep Discriminative Direct Decoders for High-dimensional Time-series Analysis by M.R. Rezaei 2020 ArXiv
6. Real-Time Point Process Filter for Multidimensional Decoding Problems Using Mixture Models by M.R. Rezaei 2020 DOI
7. Basic code D4
8. Variational auto-encoded deep gaussian processes by Z. Dai et al. 2021 ArXiv
9. Parametric Gaussian process regressors by M. Jankowiak et al. 2020 ArXiv
10. A Tutorial on Gaussian Processes by Z. Ghahramani 2010 PDF
11. An Intuitive Tutorial to Gaussian Processes Regression by J. Wang 2021 ArXiv

Riemannian Geometry and Graph Laplacian metric models

1. The goal is to create a metric behavioral forecasting model for BCI
2. The impact is to construct time-embedding metric space so that it is compatible with the generative models
3. The principle a dynamic system changes its state consequently, so we construct a metric state space that could be decomposed with one of diffusion models
4. The plan
   1. select a metric model with continuous time
   2. use Riemannian geometry and Graph-Laplacian approaches
   3. make diffusion decomposition
   4. boost decoding models with metric space

References

1. Classification of covariance matrices using a Riemannian-based kernel for BCI applications by A. Barachant et al. 2013 (Neurocomputing)
2. Multi-class Brain-Computer Interface Classification by Riemannian Geometry by A. Barachant et al.
3. Riemannian Geometry for EEG-based Brain-Computer Interfaces by M. Congedo et al.
4. Online SSVEP-based BCI using Riemannian geometry by E. K. Kalunga 2016 DOI
5. A Plug&Play P300 BCI Using Information Geometry by A. Barachant 2014 ArXiv
6. Longitudinal predictive modeling of tau progression along the structural connectome by J.Dutta et al. 2021 DOI
7. Grand: Graph neural diffusion by M.M. Bronstein et al. ICML, 2021.
8. (inspiring) The inverse problem in electroencephalography using the bidomain model of electrical activity by A.L. Rincon and S. Shimoda, 2016 DOI
9. (inverse) High-Resolution EEG Source Reconstruction with Boundary Element Fast Multipole Method, N. Makaroff et al. 2022 DOI

Data

Any data that has

1. timeline with behavioral pattern, synchronous both for source and target data
2. source time series with
   • probabilistic assumptions for diffusion probabilistic models
3. target time series with
   • behavioral pattern to make a decision

To select from

• List of datasets and tools
• Scientific lessons from a catalogue of 6674 brain recordings by A.D.J. Makin