Build · Initialization

Initializer catalog

Input and feedback initializers fill the weight matrices that project external signals into the reservoir. Each page below shows the matrix an initializer draws, lists its parameters, and gives the code to select it in ESNLayer.

• bernoulli

  ---

  Bernoulli sign initializer for input/feedback weight matrices.

• binary_balanced

  ---

  Deterministic binary balanced initializer using Walsh-Hadamard structure.

• chain_of_neurons_input

  ---

  Input initializer for chain-of-neurons reservoirs.

• chebyshev

  ---

  Chebyshev mapping initializer for deterministic chaotic initialization.

• chessboard

  ---

  Chessboard pattern initializer with alternating {-1, +1} values.

• dendrocycle_input

  ---

  Input initializer for dendro-cycle reservoirs.

• informed_init

  ---

  Informed input layer for hybrid/knowledge ESNs (RC.jl informed_init [Pathak2018]).

• logistic_mapping

  ---

  Logistic-map input initializer (RC.jl logistic_mapping [Wang2022]).

• minimal_init

  ---

  Minimal-complexity input weights (Rodan & Tiňo; RC.jl minimal_init).

• normal

  ---

  Gaussian (normal) initializer for input/feedback weight matrices.

• opposite_anchors

  ---

  Initializer that connects each input to two opposite anchors on an n-node ring.

• pseudo_diagonal

  ---

  Pseudo-diagonal initializer for structured input connections.

• random

  ---

  Random initializer for feedback/input weight matrices.

• random_binary

  ---

  Binary random initializer for input/feedback weight matrices.

• ring_window

  ---

  Deterministic windowed input initializer for ring-based topologies.

• uniform

  ---

  Bounded uniform initializer for input/feedback weight matrices.

• weighted_init

  ---

  Weighted block-diagonal random input layer (RC.jl weighted_init [Lu2017]).

• weighted_minimal

  ---

  Minimal weighted block-diagonal input layer (RC.jl weighted_minimal).

• zeros

  ---

  Initializer that sets all weights to zero.