Hebbian Memory Examples

The examples/brain_inspired/ scripts highlight how to drive HebbianTrainer with real data. They also illustrate common pre-processing steps for pattern memories.

hopfield_train.py

• Location: examples/brain_inspired/hopfield_train.py

• Scenario: Convert a handful of skimage images into pattern memories, train a Hopfield network, then recover noisy inputs.

• Key steps:

  1. preprocess_image resizes and thresholds an image to a 128×128 vector in {-1, +1}.

  2. Instantiate AmariHopfieldNetwork (synchronous updates, sign activation).

  3. Initialise HebbianTrainer(model) and call trainer.train(data_list).

  4. Corrupt each pattern by flipping ~30% of the pixels and run trainer.predict_batch.

  5. Use Matplotlib to compare train/input/output panels (saved as discrete_hopfield_train.png).

• Extensions:

  • Toggle asyn=True to observe asynchronous convergence.

  • Set normalize_by_patterns=False to keep the magnitude of original patterns.

hopfield_train_mnist.py

• Location: examples/brain_inspired/hopfield_train_mnist.py

• Scenario: Load a small selection of MNIST digits (with a series of fallbacks) and store them in a Hopfield network.

• Key steps:

  1. _load_mnist() attempts Hugging Face datasets, TorchVision, Keras, and finally scikit-learn digits.

  2. Choose one exemplar per digit class and convert with _threshold_to_pm1.

  3. Train with trainer.train(patterns).

  4. Run trainer.predict on clean held-out samples to confirm retrieval.

  5. Plot the train/input/output triads via plt.subplots.

• Extensions:

  • Introduce noisy test images to measure robustness.

  • Enable trainer.configure_progress(show_iteration_progress=True) to log energy convergence.

Custom models

• To plug in your own model, follow the BrainInspiredModel interface: expose W and s states (or override weight_attr / predict_state_attr) and provide update/energy.

• Experiment with AmariHopfieldNetwork parameters such as activation or temperature to study the effect on energy descent.