Module feature_extractor 

Differentiable feature extractors for Deep Kernel Learning.

NeuralFeatureMap is the trait an F must satisfy to be plugged into crate::deep_kernel::DeepKernel. The v0.2.0 preview ships a single implementation — MLPFeatureExtractor — a stack of DenseLayers with ReLU / Tanh / Identity activations. Future releases may add CNN / Transformer extractors; the trait is designed to keep those additions out-of-tree until they are ready.

Naming note

The trait is named NeuralFeatureMap (not FeatureExtractor) to avoid colliding with the pre-existing crate::feature_extraction::FeatureExtractor struct, which is specifically for turning tensorlogic_ir::TLExpr into numeric features. The two types coexist in the same crate and serve complementary purposes.

Structs

MLPFeatureExtractor

Multi-layer perceptron feature extractor.

Traits

NeuralFeatureMap

A differentiable map ℝ^{d_in} → ℝ^{d_out} used as the feature extractor inside a Deep Kernel.

Type Aliases

ForwardCache

Bundle returned by MLPFeatureExtractor::forward_with_cache — (final_output, per_layer_caches).

LayerCache

Per-layer cache (pre_activation, post_activation) used by the analytical backprop path in crate::deep_kernel::gradient.