Feed-Forward Networks (FFN)

Documentation status: reference — see Maturity and evidence.

A feed-forward network is a differentiable composition of linear transformations and nonlinear activations.

Basic regression shape

input
  -> linear layer
  -> activation
  -> optional hidden layers
  -> output

For classification, a final score vector can be combined with softmax and an appropriate loss such as cross-entropy.

Conceptual and dense inputs

Conceptual features can be projected into dense activations before entering conventional FFN kernels. Keep the mapping between conceptual features and dense columns explicit when interpretation matters.

Training

Trainable parameters participate in the computation graph. A training loop evaluates the forward pass, computes loss, propagates gradients, and updates parameters through the selected optimizer mechanism.

Role in logiCells

FFNs provide one numerical evaluation mechanism that can be combined with explicit conceptual structure, rules, and other neural architectures.