verskyt.utils

Utility functions and helper modules.

Module: initializers

Custom initializers for Tversky neural network parameters.

Based on empirical findings from the paper’s experiments.

uniform_init(tensor: Tensor, a: float = -1.0, b: float = 1.0)

Uniform initialization.

Paper finding: Uniform initialization led to higher convergence probability for XOR compared to normal and orthogonal.

xavier_uniform_init(tensor: Tensor, gain: float = 1.0)

Xavier/Glorot uniform initialization.

initialize_for_xor(prototypes: Tensor, features: Tensor, seed: int | None = None) → tuple[Tensor, Tensor]

Initialize parameters specifically for XOR task.

Based on the paper’s Figure 1 showing a working XOR configuration: - 2 prototypes: p0 = {}, p1 = {f0, f1} - 2 features positioned to separate XOR classes

Parameters:

• prototypes – Prototype tensor to initialize [2, 2]

• features – Feature tensor to initialize [num_features, 2]

• seed – Random seed for reproducibility

Returns:

Tuple of (initialized_prototypes, initialized_features)

smart_init(layer: Module, method: str = 'xavier_uniform', **kwargs)

Smart initialization for Tversky layers based on paper findings.

Parameters:

• layer – TverskyProjectionLayer or TverskySimilarityLayer

• method – Initialization method

• **kwargs – Additional arguments for initialization

Custom initialization strategies for Tversky Neural Network parameters.