[Paper Review] Classification of weak multi-view signals by sharing factors in a mixture of Bayesian group factor analyzers

Sami Remes, Tommi Mononen+1 more|arXiv (Cornell University)|Dec 17, 2015

Blind Source Separation Techniques5 references3 citations

TL;DR

This paper proposes a Bayesian mixture model of group factor analyzers that shares latent factors across clusters to improve classification of weak, multi-view signals—particularly in single-trial MEG data. By separating shared (common) noise components from cluster-specific (discriminative) factors, the model achieves significantly higher accuracy than baseline methods, even with very small training sets, and provides interpretable, generative reconstructions of brain responses.

ABSTRACT

We propose a novel classification model for weak signal data, building upon a recent model for Bayesian multi-view learning, Group Factor Analysis (GFA). Instead of assuming all data to come from a single GFA model, we allow latent clusters, each having a different GFA model and producing a different class distribution. We show that sharing information across the clusters, by sharing factors, increases the classification accuracy considerably; the shared factors essentially form a flexible noise model that explains away the part of data not related to classification. Motivation for the setting comes from single-trial functional brain imaging data, having a very low signal-to-noise ratio and a natural multi-view setting, with the different sensors, measurement modalities (EEG, MEG, fMRI) and possible auxiliary information as views. We demonstrate our model on a MEG dataset.

Motivation & Objective

• To address the challenge of classifying weak, low signal-to-noise ratio signals in multi-view brain imaging data, such as single-trial MEG.
• To develop a generative classification model that preserves temporal structure and enables interpretability, unlike black-box methods such as SVM or Gaussian processes.
• To improve classification performance in data-scarce settings by sharing statistical strength across clusters through shared latent factors.
• To provide a model that distinguishes between common neural noise and class-discriminative brain activity, enhancing signal detection in complex neuroimaging data.

Proposed method

• The model extends Group Factor Analysis (GFA) by modeling each class as a mixture component with its own cluster-specific factor loading matrix and shared factor loading matrix.
• Shared factors capture common, non-discriminative neural activity across classes, while cluster-specific factors model the weak, task-related signals.
• A variational Bayesian inference approach is used to jointly estimate the model parameters, including shared and cluster-specific loadings, factor scores, and class label distributions.
• The model uses a hierarchical prior structure with hyperpriors on factor precisions and a Bernoulli prior on class membership to enable robust learning with small datasets.
• Time-series structure in MEG data is preserved by modeling each sensor’s time window as a data source, with factors capturing temporal dynamics.
• The model is trained end-to-end on multi-view data, where views include different MEG sensors, modalities, or experimental conditions.

Experimental results

Research questions

• RQ1 Can shared latent factors across multiple GFA components improve classification accuracy in weak multi-view signal data?
• RQ2 How does the inclusion of shared factors compare to models with only cluster-specific factors in terms of performance on low-SNR brain imaging data?
• RQ3 Can the model effectively separate common neural noise from task-discriminative signals in single-trial MEG data?
• RQ4 Does the generative nature of the model allow for interpretable reconstruction of brain responses, such as event-related potentials (ERPs), that align with empirical averages?

Key findings

• The proposed model significantly outperforms a Group LASSO baseline, especially with small training sets, achieving high AUC even when training data is limited to just 4 samples.
• A restricted version of the model without shared factors performed poorly, demonstrating that shared factors are essential for performance improvement.
• The model successfully reconstructed class-specific ERPs that closely matched empirical averages, confirming its ability to capture discriminative neural patterns.
• The model identified brain regions associated with speaking (e.g., motor-related areas near helmet edges) and listening (e.g., bilateral auditory cortices), aligning with known neurophysiology.
• The generative model enabled interpretation of which factors contributed to class differences, with shared factors explaining non-discriminative activity and cluster-specific factors capturing task-relevant signals.