Novel Saliency Analysis for the Forward Forward Algorithm

TL;DR

This work developed a bespoke saliency algorithm specifically tailored for the ForwardForward framework, which enhances the intuitive understanding of feature importance and network decision-making, providing clear visualizations of the data features most influential in model predictions.

Abstract

Incorporating the Forward Forward algorithm into neural network training represents a transformative shift from traditional methods, introducing a dual forward mechanism that streamlines the learning process by bypassing the complexities of derivative propagation. This method is noted for its simplicity and efficiency and involves executing two forward passes the first with actual data to promote positive reinforcement, and the second with synthetically generated negative data to enable discriminative learning. Our experiments confirm that the Forward Forward algorithm is not merely an experimental novelty but a viable training strategy that competes robustly with conventional multi layer perceptron (MLP) architectures. To overcome the limitations inherent in traditional saliency techniques, which predominantly rely on gradient based methods, we developed a bespoke saliency algorithm specifically tailored for the Forward Forward framework. This innovative algorithm enhances the intuitive understanding of feature importance and network decision-making, providing clear visualizations of the data features most influential in model predictions. By leveraging this specialized saliency method, we gain deeper insights into the internal workings of the model, significantly enhancing our interpretative capabilities beyond those offered by standard approaches. Our evaluations, utilizing the MNIST and Fashion MNIST datasets, demonstrate that our method performs comparably to traditional MLP-based models.

Figures (5)

• Figure 1: MNIST images contain a black border. If we replace the first 10 pixels by one of N representations of the label. To develop positive and negative samples for the forward-forward algorithm, we change these N pixels. For a positive sample, set the tensor index corresponding to the class to 1 and all other indices to 0. This denotes the class's presence. Conversely, for a negative sample, choose a random index not matching the actual class, set this to 1, and all other indices to 0, indicating the absence of the class. This method allows the algorithm to distinguish between classes by clearly identifying which class is present (positive sample) and which is not (negative sample).
• Figure 2: Plotting the loss for each layer in the forward-forward algorithm, where each layer learns independently, starting with the first layer, then proceeding to the second
• Figure 3: Plotting the loss for each layer in the forward-forward algorithm, where each layer learns independently, starting with the first layer, then proceeding to the second
• Figure 4: The Accuracy Differential Saliency (ADS) Technique, applied to the MNIST dataset, evaluates the impact of individual pixels on model accuracy by zeroing them out with a moving filter. This process generates a difference matrix by comparing the model's performance with and without each pixel's contribution, highlighting key areas affecting decision-making. The results are visualized by plotting this matrix over the image.
• Figure 5: The Accuracy Differential Saliency (ADS) Technique, applied to the Fashion MNIST dataset, evaluates the impact of individual pixels on model accuracy by zeroing them out with a moving filter.