The core problem is the model's lack of rotational invariance, meaning it cannot recognize an object if its orientation changes. Data augmentation, specifically applying geometric transformations like rotation and flipping, directly addresses this issue. By creating augmented versions of the existing training images with various orientations, the model learns that the object's class (e.g., "panda") is independent of its orientation. This technique effectively expands the dataset's diversity without requiring the collection of new images, directly targeting the identified weakness and enhancing the model's robustness and accuracy.

B: Increasing training epochs on the original, non-diverse dataset will likely lead to overfitting on the upright images, potentially worsening the model's performance on rotated images.

C: Normalization standardizes pixel intensity and color distributions. It is a crucial preprocessing step but does not alter the geometric orientation of objects within the images.

D: While transfer learning is a powerful technique, a pre-trained model is not guaranteed to be rotationally invariant. It would still require fine-tuning on an orientation-diverse dataset, making data augmentation (A) the more direct and necessary solution.

1. Official Vendor Documentation (AWS): The Amazon SageMaker Developer Guide discusses data augmentation as a key technique to improve model accuracy by creating modified versions of existing data. It lists transformations like "randomly rotating the image" as a common practice to make the model more robust to variations in input data.

Source: Amazon SageMaker Developer Guide, "Data Augmentation" section.

2. University Courseware (Stanford): Stanford's CS231n course on Convolutional Neural Networks for Visual Recognition explicitly details data augmentation as a primary regularization technique. The notes state, "The most common data augmentation technique is horizontal flipping... Other augmentations include random crops and scales, color jittering, but also rotations..." This directly supports using rotation to handle orientation sensitivity.

Source: Stanford CS231n: Convolutional Neural Networks for Visual Recognition, Course Notes, "Training Neural Networks, Part 1", Section: Data Augmentation.

3. Peer-Reviewed Academic Publication: A survey on image data augmentation highlights that geometric transformations, including rotation and flipping, are used to "enable the model to be invariant to the position and orientation of objects in the image." This confirms that augmentation is the standard, effective approach for the described problem.

Source: Shorten, C., & Khoshgoftaar, T. M. (2019). A survey on Image Data Augmentation for Deep Learning. Journal of Big Data, 6(1), 60. Section 3.1 "Geometric transformations". DOI: https://doi.org/10.1186/s40537-019-0197-0