Skip to content

V1.md

Latest commit

 

History

History
59 lines (42 loc) · 2.66 KB

V1.md

File metadata and controls

59 lines (42 loc) · 2.66 KB
Raw

AeloRu V1 — Foundational Implementation

Adaptive Elastic Learning with Orthogonal Robust Units

V1 is the foundational, stable implementation of the AeloRu layer. It establishes the core architecture that all later versions build on: a neuroscience-inspired low-rank adapter with a conflict-driven cognitive state machine.

Detailed design (formulas, configuration reference, code examples) lives in README.md. This file is a quick entry point.

中文版：V1_zh.md

Modules Implemented

Module Principle Role
Hi-DoRA Amplitude–direction decoupling Row/column learnable amplitude vectors modulate ΔW = (m_x · m_yᵀ) ⊙ (α/r · B@A).
ReLoRA Periodic merge-and-reset Accumulates increments into external buffer W_acc; high-rank consolidation.
Hebbian-Fisher Bidirectional synaptic plasticity Hebbian trace records exploration; Fisher mask protects critical weights.
HongWen State Machine Four-phase cognitive cycle EXPLORE → RED → ANCHOR → SOLID, driven by conflict score C = 0.6·v_F + 0.4·(1−H).
Hierarchical Fisher Gradient / sparse / full snapshot Gradient (high-freq) + sparse Top-K (mid-freq) + quantized full snapshot (low-freq).
Orthogonal Penalty L_ortho = λ‖ΔWᵀW₀‖² Pushes ΔW into the left null space of W₀.
Energy Budget ‖ΔW‖_F ≤ η‖W₀‖_F Hard constraint preventing adapter from overwhelming pretrained weights.

Effective weight: W_eff = W₀ + W_acc + Gate(ΔW).

Core API

from aeloru_layer import inject_aeloru, AeloruConfig, train_aeloru_step

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)

for x, y in dataloader:
    loss, metrics = train_aeloru_step(layer, x, y_target=y, optimizer=optimizer)
    print(metrics["state"], metrics["loss_total"])

layer.save_adapter("ckpt.pt")   # incremental-only checkpoint
layer.load_adapter("ckpt.pt")

Status & Notes

  • Status: ✅ Stable foundation. All modules independently toggleable via AeloruConfig.
  • V1 uses the original 2-D parameter-level Fisher mask (d_out × d_in). The O(d²) cost was later reduced to O(d) in V2/V3 via dual 1-D HGF precision gating.
  • The conflict-detection buffer wiring was incomplete in V1 and was fixed in V3 (see V3/md/fix1.md).
  • Companion docs: README.md (English design), README_zh.md (中文), experiment/ for the hidden-state semantic analysis (P0).