Fig. 2
From: Training convolutional neural networks with the Forward–Forward Algorithm
Schematic overview of the FF-trained CNN applied to the MNIST dataset. Positive and negative samples are processed through three convolutional layers, each followed by layer normalization and ReLU activation. At every layer, the goodness function is computed using binary cross-entropy for both positive and negative samples. Final classification can then be performed either through a linear classifier or by evaluating the goodness scores across all labels.
