mnemos-hebbian 0.2.0a1

Neuromorphic AI built on Hebbian learning. No backprop. No gradients. No loss functions.

Mnemos

Neuromorphic AI built on Hebbian learning. No backprop. No gradients. No loss functions.

⚠️ Alpha — APIs may change. Benchmarks are real and reproducible.

pip install mnemos

---

Results

Bearing fault detection — real industrial data

Metric	Mnemos	Standard 3-sigma	Deep learning (GPU)
F1 score	0.876 ± 0.018	0.000	0.82–0.88
Recall	1.000	0.000	~0.85
Warning time	6.2 hours	0 hours	unknown
False alarms/day	0.0	—	unknown
GPU needed	No	No	Yes
Labels needed	No	No	Yes

Tested on NASA IMS and CWRU — two standard benchmarks used in published papers. Same code, no modifications between datasets. CWRU F1: 0.998.

The standard industrial method (3-sigma RMS threshold) gives zero warning on this failure mode because the bearing degrades gradually. Mnemos catches it 6.2 hours early, automatically, with no manual tuning.

detector = mnemos.AnomalyDetector()
detector.fit(normal_vibration_data)    # learns normal in <1s, no labels
scores = detector.score(new_data)      # higher = more anomalous

---

AdaptiveHead — plug into any frozen backbone, replace fine-tuning

Metric	Mnemos	PyTorch backprop
CIFAR-10 accuracy	82.2%	85.2%
Compute cost	258M ops	234,000M ops
Compute savings	905x less	baseline
Add new class	0.3s, no retraining	full retrain
Forgetting (new classes)	−18%	N/A

head = mnemos.AdaptiveHead(n_proto=10)
head.fit(train_features, train_labels)       # 1.1s, no backprop
head.add_class("new_category", new_features) # 0.3s, no retraining
predictions = head.predict(test_features)

---

AnomalyDetector — beats sklearn IsolationForest on 5/5 benchmarks

Dataset	Mnemos	IsolationForest
Gaussian 10D	0.840 ± 0.063	0.621
High-dim 50D	1.000 ± 0.000	0.893
Noisy 10D	0.952 ± 0.025	0.901
Multi-cluster	0.977 ± 0.016	0.885
Sparse 1%	0.795 ± 0.055	0.512

15/15 stress tests passed (drift, adversarial, stability).

---

ContinualLearner — learns new tasks without forgetting

Method	Split-MNIST accuracy	Uses gradients
Mnemos	97.4%	No
EWC (DeepMind)	65.5%	Yes

---

What is Mnemos?

Every weight update in Mnemos follows one rule:

dw = f(pre, post)

Pre-synaptic activity times post-synaptic activity. Local only. No global error signal. No backward pass. This is how biological neurons learn.

The key contribution is a specificity penalty that prevents winner-take-all collapse in competitive Hebbian networks:

effective_similarity = raw_similarity - mean_similarity - threshold

Without this, competitive learning degenerates — all neurons converge to the same representation. With it, the network maintains diverse, class-specific prototypes. This appears in every competitive layer in Mnemos.

AdaptiveHead additionally uses LVQ2.1 with spherical tangent geometry for boundary refinement. The standard push/pull update is geometrically wrong on normalized vectors — it pushes toward the origin rather than away from the confuser class. Fixing this was the largest single accuracy improvement.

---

Installation

pip install mnemos

From source:

git clone https://github.com/theGcmd/mnemos.git
cd mnemos
pip install -e .

---

Modules

Module	What it does
AnomalyDetector	Learn normal, flag anomalies. No labels needed.
AdaptiveHead	Plug into any frozen PyTorch backbone. Replace fine-tuning.
ContinualLearner	Learn new tasks without forgetting old ones.
HebbianFilters	Competitive feature learning with specificity penalty.
HebbianMemory	Associative knowledge storage in weight matrices.
PrototypeBridge	Multi-prototype pattern recognition.
Brain	Full perception → recognition → reasoning loop.
StreamProcessor	Real-time stream monitoring with drift detection.

---

How AdaptiveHead works

Frozen backbone → 512-dim features → Hebbian clustering → LVQ2.1 refinement → Nearest prototype

1. Hebbian clustering — competitive k-means using only local updates
2. LVQ2.1 refinement — winner pulled toward sample, confuser pushed away, spherical tangent geometry
3. Prediction — nearest prototype by cosine similarity

No backward(). No optimizer state. No gradient storage.

---

Honest limitations

• AdaptiveHead has a 3% accuracy gap vs backprop on CIFAR-10
• add_class() has ~18% forgetting when new classes have high feature overlap with existing ones. This matches published SOTA for gradient-free incremental learning.
• Bearing fault detection tested on two standard benchmarks — real-world deployment may surface additional challenges (variable loading, sensor noise, temperature drift)

---

License

Free for research and non-commercial use. Commercial use requires a license from the author. See LICENSE for details.

---

Citation

@software{mnemos2026,
  author = {Gausepohl, Gustav},
  title  = {Mnemos: Neuromorphic AI Built on Hebbian Learning},
  year   = {2026},
  url    = {https://github.com/theGcmd/mnemos}
}

---

Author

Gustav Gausepohl — independent AI researcher, age 14, UK. thegcmd.github.io · [e-mail]

"Neurons that fire together, wire together." — Donald Hebb, 1949