Hierarchical Resonance Foundation Model
A hierarchical System 1 model with per-example resonance, trained feedback, and bounded adaptation.
Overview
The HierarchicalResonanceFoundationModel organizes ResonantGSER layers into a feedback-driven sequence model. During resonance prediction, every example carries its own top-down target through the hierarchy. Feedback training and test-time adaptation remain separate, measurable operations with clear reset semantics.
Usage
foundation-model.py
from gpbacay_arcane import HierarchicalResonanceFoundationModelmodel = HierarchicalResonanceFoundationModel( vocab_size=5000, seq_len=32, hidden_dim=128, num_resonance_levels=4, enable_inference_plasticity=True, use_hierarchical_homeostasis=True, homeostasis_config={ "homeostasis_mode": "hybrid", "homeostatic_window_size": 32, "slow_bcm_tau": 4000.0, "adaptive_target_rate": 0.05, "resonance_coupling": 0.5, "hierarchy_target_slope": 0.1, },)model.build_model()model.compile_model()# Ordinary supervised trainingmodel.model.fit(X_train, y_train)# Train task weights and feedback projections end to endmetrics = model.train_feedback_step( X_train[:32], labels=y_train[:32], alignment_weight=0.1)# Refine the current batch with isolated per-example resonance statepredictions = model.predict_with_resonance(X_test, resonance_cycles=4)# Optional bounded adaptation; episodic=True restores the checkpoint afterwardresult = model.adapt_at_inference( X_test, variables="all", objective="entropy", steps=2, max_update_norm=1e-3, episodic=True,)Key Features
- Per-example resonance: explicit batch-shaped targets avoid co-batching contamination during current-input refinement.
- Trainable feedback: alignment and reconstruction losses provide a gradient path to feedback projections.
- Bounded adaptation: optional entropy or consistency objectives support snapshots, update limits, rollback, and episodic reset.
- Temporal Coherence: Distills temporal dynamics into coherence vectors.
- Attention Fusion: Aggregates multi-pathway information with self-attention.
Choose the resonance depth with a controlled ablation of the same checkpoint at 0, 1, 2, 4, and 8 cycles. This reveals the best accuracy, calibration, and latency trade-off for your task.