src.canns.trainer.oja

Oja’s normalized Hebbian learning trainer.

Classes

OjaTrainer

Oja's normalized Hebbian learning trainer.

Module Contents

class src.canns.trainer.oja.OjaTrainer(model, learning_rate=0.01, normalize_weights=True, weight_attr='W', compiled=True, **kwargs)

Bases: src.canns.trainer._base.Trainer

Oja’s normalized Hebbian learning trainer.

Oja’s rule stabilizes pure Hebbian growth by introducing a weight-dependent normalization term, enabling single-neuron principal component extraction without unbounded weight magnitudes.

Learning Rule:

ΔW_ij = η * (y_i * x_j - y_i^2 * W_ij)

where:

• W_ij is the weight from input j to output i

• x_j is the input activity

• y_i is the output activity (y = W @ x)

• η is the learning rate

The rule can be rewritten as:

ΔW = η * (y @ x^T - diag(y^2) @ W)

This naturally leads to weight normalization and PCA extraction.

Reference:

Oja, E. (1982). Simplified neuron model as a principal component analyzer. Journal of Mathematical Biology, 15(3), 267-273.

Initialize Oja trainer.

Parameters:

• model (src.canns.models.brain_inspired.BrainInspiredModel) – The model to train (typically LinearLayer)

• learning_rate (float) – Learning rate η for weight updates

• normalize_weights (bool) – Whether to normalize weights to unit norm after each update

• weight_attr (str) – Name of model attribute holding the connection weights

• compiled (bool) – Whether to use JIT-compiled training loop (default: True)

• **kwargs – Additional arguments passed to parent Trainer

predict(pattern, *args, **kwargs)

Predict output for a single input pattern.

Parameters:

pattern – Input pattern of shape (input_size,)

Returns:

Output pattern of shape (output_size,)

train(train_data)

Train the model using Oja’s rule.

Parameters:

train_data (collections.abc.Iterable) – Iterable of input patterns (each of shape (input_size,))

compiled = True

learning_rate = 0.01

normalize_weights = True

weight_attr = 'W'