YAML Config Glossary

YAML Config Glossary#

This glossary describes all the optional parameters that can be specified in the YAML configuration file when building a neural network using the EIANN library.

1. Layer Config#

Defines the structure and behavior of neuron populations in each layer.

Structure#

layer_config:
  LayerName:
    PopulationName:
      size: int
      activation: str
      activation_kwargs: dict
      include_bias: bool
      bias_init: str
      bias_init_args: tuple
      bias_bounds: [float, float]
      bias_learning_rule: str
      bias_learning_rule_kwargs: dict
      custom_update: str
      custom_update_kwargs: dict
      output_pop: bool

Parameters#

  • size: Number of neurons in the population.

  • activation: Name of the nonlinearity function (e.g. “relu”, “sigmoid”).

  • activation_kwargs: Keyword arguments passed to the activation function.

  • include_bias: Whether to include a trainable bias term.

  • bias_init: Callable used to initialize the bias.

  • bias_init_args: Arguments passed to the bias initializer.

  • bias_bounds: Tuple of min/max values for clipping the bias.

  • bias_learning_rule: Callable name for the learning rule applied to the bias.

  • bias_learning_rule_kwargs: Additional parameters for the bias learning rule.

  • output_pop: Marks this population as the output target for loss calculation.

2. Projection Config#

Defines all projections (i.e., weighted connections) between neuron populations.

Structure#

projection_config:

  PostLayer.PostPopulation:
    PreLayer.PrePopulation:
      weight_init: str
      weight_init_args: tuple
      weight_constraint: str
      weight_constraint_kwargs: dict
      weight_bounds: [float, float]
      direction: str
      update_phase: str
      compartment: str
      learning_rule: str
      learning_rule_kwargs: dict

Parameters#

  • weight_init: Name of the weight initializer function (e.g. “half_kaiming”).

  • weight_init_args: Arguments passed to the initializer. If a single number is provided, it is used as scaling factor when initializing weights.

  • weight_constraint: Optional constraint on the weight matrix. Options: normalize_weight, clone_weight, no_autapses

  • weight_constraint_kwargs: Parameters for the weight constraint.

  • weight_bounds: Tuple of [min, max] values to clip weights. “Null” means no clipping.

  • direction: Forward or Recurrent connection type. Recurrent means the connection uses activities from the previous time step. Options: ‘forward’,’F’, ‘recurrent’,’R’

  • update_phase: Informs the learning rule to determine which phase of activations to use for weight updates. Options: ‘forward’ (default option),’F’, ‘backward’,’B’, ‘all’,’A’

  • compartment: Specifies target compartment (‘soma’ or ‘dendrite’).

  • learning_rule: Name of the function used to update weights (e.g. “Backprop”, “BCM”, “Hebb_WeightNorm”).

  • learning_rule_kwargs: Keyword arguments passed to the learning rule.

3. Training kwargs#

Global settings for running the training loop and initializing the network.

Structure#

training_kwargs:
  learning_rate: float
  optimizer: str
  optimizer_kwargs: dict
  criterion: str
  criterion_kwargs: dict
  seed: int
  device: str
  tau: int
  forward_steps: int
  backward_steps: int
  verbose: bool

Parameters#

  • learning_rate: Global fallback learning rate for all trainable parameters.

  • optimizer: Name of optimizer class (default: SGD).

  • optimizer_kwargs: Arguments passed to the optimizer (e.g. momentum).

  • criterion: Loss function name (e.g. “MSELoss”).

  • criterion_kwargs: Additional arguments passed to the loss function.

  • seed: Random seed for reproducibility.

  • device: Hardware device (“cpu” or “cuda”). [Note: This functionality has not been fully tested for “cuda”]

  • tau: Decay time constant for unit leak dynamics. tau=1 represents complete decay in each step (i.e. no persistent state)

  • forward_steps: Number of simulation steps per forward pass.

  • backward_steps: Number of steps for backprop or backward phase (if applicable). In the case of Backprop, results in truncated BPTT.

  • verbose: Enables logging or debug outputs during training.

Contents