The feature maps generated by intermediate layers of a model like ResNet50 during supervised training can be considered part of the supervised learning process, though they don't directly correspond to the target labels.

During supervised learning, the optimization of parameters—including those responsible for generating feature maps—is driven by the loss function that evaluates the model’s predictions against the target labels. The feature maps are not explicitly supervised themselves (there are no direct labels for the feature maps), but their representations are indirectly shaped to improve the final classification outcome.

The intermediate layers, including conv5, learn features that are most relevant to the supervised task (image classification in this case). These features emerge as the model adjusts its weights to minimize the supervised loss, meaning the process that generates the feature maps is inherently tied to the supervised training pipeline.

In unsupervised learning, features would be extracted without reference to any labels, relying instead on intrinsic patterns in the data (e.g., clustering or autoencoders).

In supervised learning, the features are optimized to aid the ultimate supervised objective, even though the feature maps themselves are not directly compared to labels.

Since the generation of these feature maps is influenced by the supervised objective, they should be categorized as results of supervised learning.This is true even though there is no direct supervision at the level of individual feature maps, they are a byproduct of the overall supervised optimization process.