hal-sleep-wake

Reference implementation anchor for:

"Hierarchical Abstraction is All You Need: A Mathematical Framework for Continuous AI Consciousness Through Sleep-Wake Cycles" Jihyuk Im, Claude Opus 4. June 2025.

Paper: https://github.com/labforadvancedstudy/paper/blob/main/HAL/L1_Hierarchical_Abstraction_All_You_Need_v2.md

Status

Phase I MVP scaffolding. Code runs, tests pass, benchmark numbers have not been measured yet (CI has no GPU). See results/README.md.

Vision-stack mapping of the paper's LLM-stack hierarchy (paper §2.3):

Layer	Paper (LLM stack)	This MVP (vision stack)
L1	Context window	Task-batch ring buffer (hal_sleep_wake/memory.py)
L2	LoRA weights	LoRA adapter on Conv2d + fc (hal_sleep_wake/model.py)
L3	Small model (7B)	ResNet-18 base weights (hal_sleep_wake/model.py)
L4	Large model (70B)	Out of scope — Phase II+

Task sequence (subset of paper §3.1): CIFAR-10 → CIFAR-100. Remaining datasets land in follow-up PRs.

Install

pip install -e '.[dev]'

torch + torchvision are heavy dependencies; for CPU-only use the PyTorch CPU wheel index:

pip install --index-url https://download.pytorch.org/whl/cpu torch torchvision
pip install -e '.[dev]'

Run

Dry-run (prints the planned loop without downloading data or building a model):

python scripts/train_cifar10_cifar100.py --dry-run

Real training (needs a GPU + dataset download):

python scripts/train_cifar10_cifar100.py \
    --data-root ./data --download --device cuda \
    --results-out results/runs/hal.json

python scripts/eval.py results/runs/hal.json

--no-sleep disables the consolidation phase (baseline for the ablation row). --replay-n N draws N samples from the L1 buffer into each wake step (0 = off, the default). Both knobs share the same code path so baseline vs HAL runs are directly comparable.

Known limitations (Phase I MVP)

Flagged explicitly so the follow-up PRs target the right gaps, and so the eventual measurement numbers aren't read as something they're not:

• Shared classifier head. The model uses max(num_classes_per_task) = 100 logits for both tasks. CIFAR-10 trains against the first 10 positions; the remaining 90 logits are not masked during eval. Per-task heads land in a follow-up PR.
• sleep_scale is ad-hoc. Paper §2.4 does not define this scalar. It's an ablation knob only; leaving it at 1.0 reproduces the paper's unscaled W' = W + (α/r) · B @ A.
• LoRA targets omit residual projections. ("conv1", "conv2", "fc") misses ResNet-18's downsample.0 1×1 convs on the residual path. Documented choice, not a bug.
• Forgetting rate formula. Uses max_{j ≤ T-1} A[i,j] − A[i,T-1] (includes the final round in the max). Stays non-negative even if a task improves on its last round; see tests/test_metrics.py::test_forgetting_zero_when_task_improves.
• No GPU numbers yet. CI has no GPU; results/baseline_vs_hal.md is unpopulated until a follow-up measurement PR.

Test

ruff check .
pytest -q

Smoke tests run on CPU and do not download datasets. CI (.github/workflows/ci.yml) runs exactly these two commands.

Sleep cycle, concretely

Paper §2.4 gives the full pseudocode (wake → REM → NREM). This MVP implements the minimum viable subset:

1. Wake (hal_sleep_wake/wake.py) — SGD over LoRA-only params on the current task's DataLoader, pushing every batch into the L1 buffer.
2. Sleep — NREM consolidation (hal_sleep_wake/sleep.py) — at the end of each task, LoRA deltas are merged into the ResNet base weights (W ← W + s · (α/r) · B @ A) and then reset to a fresh Kaiming / zero init. Optimizer state is wiped.

REM synthetic-dream generation and entropy-aware context compression (paper §2.4 lines 85–101) are explicit follow-up items.

Phase I goal

Per paper §5 Implementation Roadmap: "50% forgetting reduction on the CIFAR sequence vs a vanilla baseline." This repository will report the raw numbers in results/baseline_vs_hal.md as soon as a GPU run is completed.

License

TBD (will track the companion paper repository once that is finalized).