RBM

RBM implements a rotation-invariant extension of Restricted Boltzmann Machines to learn rotation-invariant feature representations from 2D images by factorizing rotational variations for bioinformatics image analysis.

Key Features:

• Rotation-Invariant Feature Learning: Learns features invariant to rotation by incorporating a mechanism to factorize rotational variations within the model architecture.
• Unsupervised Framework: Trains without labeled data using an unsupervised RBM-based framework.
• Orientation Inference: Infers an orientation for each input image during training based on reconstruction error to support invariant representation.
• Regularization via Kullback-Leibler Divergence: Employs Kullback-Leibler divergence regularization to stabilize and ensure consistency of learned features during training.
• Quantitative Evaluation with γ-Score: Uses the γ-score metric to quantify the degree of invariance achieved by the model.
• Empirical Benchmark Performance: Demonstrated improved performance over existing RBM approaches using three benchmark datasets and test accuracy measured with a Support Vector Machine (SVM) classifier.

Scientific Applications:

• Medical imaging: Provides rotation-invariant features for image-based diagnostic analyses where rotations are common.
• Molecular biology (protein structure analysis): Applies to 2D representations in protein structure analysis to reduce sensitivity to rotational variation.
• Bioinformatics image-based classification: Enhances downstream classification tasks in bioinformatics that are affected by rotational nuisance factors.

Methodology:

Unsupervised training of an RBM extension on 2D image datasets that infers per-image orientations from reconstruction error and factorizes rotational variations by adjusting internal parameters, with training regularized by Kullback-Leibler divergence and invariance evaluated via the γ-score and SVM classifier accuracy on benchmark datasets.