SGM Continual Learning

Binary locking: 1.0000x retention, zero gradient interference, three lines of code.

The Primitive

if lock_mask[i] == True:
    delta[i] = 0  # This dimension CANNOT change.

When a parameter converges, it locks. Gradient becomes zero. That coordinate is removed from the computation graph entirely. The next task trains on remaining free dimensions. Forgetting cannot happen because locked coordinates cannot move. Not approximately. Exactly.

Results

Test	Tasks	Locked	Baseline	Script
Sequential retention	1000	1.0000x	12.79x degradation	tests/ultimate_stress.py
Gradient interference	--	0.000000	EWC: 6.14	tests/test_interference.py
Academic validation	9 tests	9/9 PASS	--	experiments/academic_validation.py
Hard-lock checksum	5 tasks	Byte-exact	--	experiments/masked_forward_isolation.py
Real MNIST (PyTorch)	5	0.992-1.000	0.996-1.000	experiments/mnist_cifar_combined.py
Real CIFAR-10 (PyTorch)	5	0.827-0.932	0.823-0.930	experiments/mnist_cifar_combined.py
Parameter scaling	10M	1.00x	1.35x	tests/extreme_scale.py
12-layer transformer	12 layers	1.00x all	1.00-5.04x	tests/extreme_scale.py
8-bit quantization	--	No degradation	--	tests/quantization.py
Adversarial oscillation	100 osc.	1.0000x	--	tests/ultimate_stress.py

See VALIDATED_RESULTS.md for complete tables.

Experiments

Script	What It Tests
experiments/academic_validation.py	9 formal tests: invariance, isolation, scale, orthogonality
experiments/mnist_cifar_combined.py	Real MNIST + CIFAR-10 on GPU with PyTorch
experiments/split_mnist.py	Sequential digit pairs with retention tracking
experiments/masked_forward_isolation.py	SHA-256 checksum proof of exact parameter preservation
experiments/plasticity_amplification.py	Per-dimension learning rate amplification on free dims
experiments/real_benchmarks.py	Extended real-data benchmarks

Tests (14 scripts)

Script	What It Verifies
tests/extreme_scale.py	500 tasks, 10M params, 12-layer transformer, adversarial
tests/ultimate_stress.py	1000 tasks, oscillating targets, edge cases
tests/realworld_stress.py	Task overlap, cross-domain transfer, perturbation
tests/dense_overlap.py	Full feature overlap with sklearn digits
tests/model_scaling.py	NN + transformer scaling, inference sparsity
tests/quantization.py	8-bit quantization survival
tests/hierarchical_task_sim.py	CIFAR-like and NLP-like hierarchical tasks
tests/sklearn_digits.py	Real sklearn digits, 64-32-16-10 NN

Quick Start

git clone https://github.com/ACD421/sgm-continual-learning.git
cd sgm-continual-learning
pip install -r requirements.txt

# Core proof
python -c "
from core.sgm_core_primitives import SGMWithLocking, SparseRegionTask
import numpy as np
np.random.seed(42)
model = SGMWithLocking(dim=100)
task1 = SparseRegionTask(100, (0.0, 0.5), seed=1)
task2 = SparseRegionTask(100, (0.5, 1.0), seed=2)
model.step(task1, 1000)
t1_loss = task1.loss(model.best_x)
model.step(task2, 1000)
print(f'Locked dims: {int(np.sum(model.lock > 0.5))}/100')
print(f'Task 1 retention: {task1.loss(model.best_x)/t1_loss:.4f}x')
"

# Full test suite
python tests/extreme_scale.py --full
python experiments/academic_validation.py
python experiments/mnist_cifar_combined.py  # Needs GPU

Dependencies

numpy>=1.24.0
scipy>=1.10.0
torch>=2.0.0          # For MNIST/CIFAR experiments
torchvision>=0.15.0   # For real data loading
scikit-learn>=1.3.0   # For sklearn digits tests

Author

Andrew Dorman -- Independent AI researcher, Southlake TX

• GitHub: ACD421

License

Proprietary. See LICENSE for terms.