Table 5 Comparison of both the models

Feature Representation

• Uses scalar outputs where each neuron represents the presence of a feature.

• Capsules produce vector outputs that represent both feature presence and spatial properties.

Dynamic Routing

• CNN layers have fixed connections without

dynamic communication.

• Utilizes dynamic routing by agreement, allowing flexible communication between capsules based on feature agreement.

Handling Spatial Hierarchies

• It has difficulty in capturing part-whole relationships in complex structures.

• Effectively captures part-whole relationships, improving its ability to handle spatial hierarchies and transformations.

Robustness to Transformations

• Less robust to variations like rotation, scaling, or translation.

• More resistant to such transformations due to capsule vectors and dynamic routing.

Training Complexity

• Simpler and faster to train with lower computational requirements.

• More complex and computationally demanding due to its capsule architecture and routing process.

Overfitting and Generalization

• Prone to overfitting, often requiring dropout and other regularization methods.

• Better generalization and less prone to overfitting, thanks to richer feature representations and routing mechanisms.

Common Application Areas

• Extensively used for image classification, object detection, and related tasks.

• Suitable for tasks requiring strong spatial understanding, such as detailed image recognition and segmentation.

Interpretability

• Limited interpretability, as neurons only capture whether a feature is present or not.

• Provides higher interpretability by capturing both feature presence and spatial characteristics within capsules.