OSLNet

OSLNet enhances classification of small-sample datasets by introducing an Orthogonal Softmax Layer (OSL) that enforces orthogonal weight vectors to improve discriminative feature learning and reduce overfitting.

Key Features:

• Orthogonal Softmax Layer (OSL): Enforces orthogonality of weight vectors in the classification layer during training and testing to reduce feature redundancy and enlarge decision margins.
• Reduced Rademacher Complexity: Lowers the network Rademacher complexity to 1/K (where K is the number of classes) compared to traditional fully connected classification layers, yielding tighter generalization error bounds.
• Enhanced Generalization: Improves generalization for small-sample datasets and demonstrates superior performance compared to existing methods across benchmark datasets.

Scientific Applications:

• Biomedical Data Analysis: Applicable to genomics and proteomics classification tasks where large datasets are difficult to obtain.
• Rare Disease Diagnosis: Suited for classification problems involving limited patient samples common in rare disease studies.

Methodology:

Constructs a deep neural network with multiple nonlinear layers and incorporates an Orthogonal Softmax Layer that maintains orthogonal weight vectors during training and testing to reduce overfitting and improve generalization.