src.canns.models.brain_inspired.linear

Generic linear layer for brain-inspired learning algorithms.

Classes

LinearLayer

Generic linear feedforward layer supporting multiple brain-inspired learning rules.

Module Contents

class src.canns.models.brain_inspired.linear.LinearLayer(input_size, output_size, use_bcm_threshold=False, threshold_tau=100.0, **kwargs)

Bases: src.canns.models.brain_inspired._base.BrainInspiredModel

Generic linear feedforward layer supporting multiple brain-inspired learning rules.

This model provides a simple linear transformation with optional sliding threshold for BCM-style plasticity. It can be used with various trainers: - OjaTrainer: Normalized Hebbian learning for PCA - BCMTrainer: Sliding threshold plasticity (requires use_bcm_threshold=True) - HebbianTrainer: Standard Hebbian learning

Computation:

y = W @ x

where W is the weight matrix, x is the input, and y is the output.

For BCM learning, an optional sliding threshold θ tracks output activity:

θ ← θ + (1/τ) * (y² - θ)

References

• Oja (1982): Simplified neuron model as a principal component analyzer

• Bienenstock et al. (1982): Theory for the development of neuron selectivity

Initialize the linear layer.

Parameters:

• input_size (int) – Dimensionality of input vectors

• output_size (int) – Number of output neurons (features to extract)

• use_bcm_threshold (bool) – Whether to maintain sliding threshold for BCM learning

• threshold_tau (float) – Time constant for threshold sliding average (only used if use_bcm_threshold=True)

• **kwargs – Additional arguments passed to parent class

forward(x)

Forward pass through the layer.

Parameters:

x (jax.numpy.ndarray) – Input vector of shape (input_size,)

Returns:

Output vector of shape (output_size,)

Return type:

jax.numpy.ndarray

resize(input_size, output_size=None, preserve_submatrix=True)

Resize layer dimensions.

Parameters:

• input_size (int) – New input dimension

• output_size (int | None) – New output dimension (if None, keep current)

• preserve_submatrix (bool) – Whether to preserve existing weights

update(prev_energy)

Update method for trainer compatibility (no-op for feedforward layer).

update_threshold()

Update the sliding threshold based on recent activity (BCM only).

This method should be called by BCMTrainer after each forward pass. Updates θ using: θ ← θ + (1/τ) * (y² - θ)

W

property energy: float

Energy for trainer compatibility (0 for feedforward layer).

input_size

output_size

property predict_state_attr: str

Name of output state for prediction.

threshold_tau = 100.0

use_bcm_threshold = False

property weight_attr: str

Name of weight parameter for generic training.

x

y