Struct ProgressiveNetworks 

pub struct ProgressiveNetworks<T: Float + Debug + Send + Sync + 'static> { /* private fields */ }

Progressive Networks for continual learning.

Each new task gets its own column (set of layers). Previous columns are frozen and lateral connections from frozen columns feed into the new column, enabling knowledge transfer without forgetting.

Implementations

impl<T: Float + Debug + Send + Sync + 'static> ProgressiveNetworks<T>

pub fn new(hidden_sizes: Vec<usize>) -> Self

Create a new progressive network with the given hidden layer sizes.

No columns are created initially; call add_task_column for each task.

pub fn add_task_column( &mut self, input_size: usize, output_size: usize, ) -> Result<usize>

Add a new column for a new task.

input_size - dimensionality of the input features. output_size - dimensionality of the output.

Returns the column index (task id). All previously existing columns are frozen.

pub fn forward(&self, input: &Array1<T>, task_id: usize) -> Result<Array1<T>>

Forward pass through the network for a specific task.

For columns beyond the first, lateral connections from all previous columns are added at each hidden layer.

pub fn forward_with_laterals( &self, input: &Array1<T>, task_id: usize, ) -> Result<(Array1<T>, Vec<Array1<T>>)>

Forward pass that also returns intermediate activations from all columns.

Returns (final_output, all_intermediate_activations) where all_intermediate_activations[col][layer] is the activation at that layer of that column.

pub fn freeze_column(&mut self, column_id: usize) -> Result<()>

Freeze a specific column (mark as non-trainable).

pub fn num_columns(&self) -> usize

Return the number of columns (one per task).

pub fn get_column_parameters( &self, column_id: usize, ) -> Result<&NetworkColumn<T>>

Get a reference to a specific column’s parameters.