AeloRu

Adaptive Elastic Learning with Orthogonal Robust Units

Beyond Low-Rank: Dynamic Plasticity through Amplitude-Direction Decoupling

Paper: AeloRu: Adaptive Elastic Learning with Orthogonal Robust Units

中文版：README_zh.md

AeloRu is a neuroscience-inspired low-rank adaptation (PEFT) framework that integrates multiple synergistic cognitive learning mechanisms while staying memory-efficient for consumer-grade GPUs. Rather than treating fine-tuning purely as a statistical-optimization problem, AeloRu borrows principles from biological learning — adaptive synaptic plasticity, phasic cognitive cycles, offline sleep consolidation, and volatility-adaptive learning rates — and casts them into a unified low-rank adapter architecture.

Core Philosophy: Pretrained weights are sacrosanct and inviolable; all learning outcomes are accumulated externally. Effective weight: W_eff = W_0 + W_acc + Gate(ΔW).

Core Research Question: How can integrating Hebbian Learning principles with LoRA-based parameter-efficient fine-tuning mitigate catastrophic forgetting during long-term sequential task learning, while enhancing adaptation speed in short-term scenarios?

Evaluated on Qwen2.5-0.5B under an extreme concept-drift protocol (SST-2 → MRPC → QNLI → RTE → CoLA, never revisited) at an equal parameter budget:

• Catastrophic forgetting reduced from LoRA's 0.0538 to 0.0076 (−85.9%)
• Backward Transfer (BWT) flipped from −0.0471 to +0.0011
• Final accuracy improved by +0.83 pp
• Cost: only +22.1% extra training time
• Geometry: inter-dimension correlation −46% ~ −65%, condition number −45% ~ −69%, rank-budget utilization 43% ~ 61% → 71% ~ 85%, task selectivity −78% ~ −86%

The gains stem from eliminating low-rank subspace redundancy rather than regularization alone. The paper further decomposes feature coupling into functional vs. redundant parts, runs a PEM dose–response experiment, and documents 13 blocking defects (and fixes) found while porting cognitive-science modules onto a pretrained Transformer. (Full details in paper.md.)

• V1 — Core AeloRu layer, HongWen state machine, ReLoRA merge, Hebbian-Fisher → V1.md · V1/README.md
• V2 — Dual 1-D HGF plasticity gating, closed-form CE gradient, PEM/DLAM integration → V2.md · V2/HGF_CE.md
• V3 — Full benchmark suite (ex1–ex4), PEM dose–response, subspace-geometry analysis, 13 defect fixes → V3.md · V3/md/

AeloRu/
├── V1/                 # Core implementation: aeloru_layer.py, HongWen state machine
├── V2/                 # Dual 1-D HGF gating, closed-form gradients (experiment_1.py)
├── V3/                 # Benchmark experiments ex1–ex4, subspace-geometry analysis
│   ├── md/             # Experiment notes (ex1, ex2, ex3, fix1…)
│   └── try_ex/         # Validation & probing scripts
├── experiment/         # Hidden-state semantic analysis (P0): analyze_hidden_states*.py
├── paper/              # Paper draft (paper.md), figures, experiment data (ex1–ex4)
├── tests/              # Unit tests (test_aeloru_logging.py)
├── logs/               # Training logs
├── output/ plots/ visualization/   # Generated artifacts
└── README.md

from V1.aeloru_layer import inject_aeloru, AeloruConfig

config = AeloruConfig(
    r=8,
    lora_alpha=4.0,
    use_hidora=True,
    use_relora=True,
    use_hongwen=True,
)
model = inject_aeloru(model, target_names=["q_proj", "v_proj"], cfg=config)

See V1/README.md for the full configuration reference (AeloruConfig), the single-step training loop (train_aeloru_step), and adapter checkpoint save/load.

1. Hebbian-RL hybrid learning for real-time adaptive plasticity
2. Adapter-level asynchrony enabling zero-latency serving during continuous learning
3. Dual 1-D HGF plasticity gating — O(d²) → O(d) Fisher gating for consumer-grade GPUs
4. DLAM spectral sleep for offline consolidation and anti-forgetting

MIT License - see LICENSE for details.