Learned Predicates

class jlnn.nn.predicates.FixedPredicate(*args: Any, **kwargs: Any)

Bases: Module

Non-trainable predicate that returns the input interval unchanged.

This module serves as an identity transformation for truth intervals, preserving the original lower and upper bounds. It is primarily utilized for crisp boolean logic scenarios where fixed truth values are required without neural weight updates.

None

class jlnn.nn.predicates.LearnedPredicate(*args: Any, **kwargs: Any)

Bases: Module

Grounding layer that transforms real-valued input data into truth intervals [L, U].

In Logical Neural Networks (LNN), predicates act as the interface between numeric data and logical formulas. This class learns the semantic mapping (grounding) by adjusting slopes and offsets of activation functions to produce fuzzy truth values.

slope_l

Steepness of the lower bound activation.

Type:

nnx.Param

offset_l

Horizontal shift for the lower bound activation.

Type:

nnx.Param

slope_u

Steepness of the upper bound activation.

Type:

nnx.Param

offset_u

Horizontal shift for the upper bound activation.

Type:

nnx.Param