H-OS-ELM¶

H-OS-ELM is an online model that stacks learned ELM-AE feature layers and trains the final head with RlsSolver.

Architecture¶

flowchart TD
    X[Input X] --> AE1[ELM-AE layer 1]
    AE1 --> F1[Learned features 1]
    F1 --> AE2[ELM-AE layer 2]
    AE2 --> F2[Final features]
    F2 --> RLS[RlsSolver]
    RLS --> Y[Predictions]

API¶

feature_elm::HierarchicalOsElm<double> model(
    numInputs,
    {64, 32},
    feature_elm::ActivationFunction::kSigmoid,
    feature_elm::Backend::kCpu,
    feature_elm::RlsOptions<double>{},
    1e-6,
    42u);

model.initialize(trainData, trainTargets, numSamples, numOutputs);
model.update(newData, newTargets, newSamples);
auto predictions = model.predictBatch(testData, testSamples);

Behavior¶

• initialize fits the ELM-AE stack greedily on the initialization data.
• update consumes the learned features from new chunks and updates only the online RLS head.
• featureStack() exposes the learned stack for inspection and tests.
• ReOS-ELM, FOS-ELM, and OS-CELM toggles are attached through RlsOptions.

When to use¶

• Data arrives online and learned feature extraction is needed.
• A single online additive layer is not expressive enough.
• You need to compare online learned features against batch ML-ELM.

When not to use¶

• Full-batch training is acceptable and simpler: use MlElm.
• You only need a random additive baseline: use OsElm.