NCA-AIIO Question 84
Select 3You are tasked with evaluating the performance of a deep learning model for image classification using NVIDIA GPUs. During the evaluation, you notice that the model performs significantly worse on a subset of the validation dataset that contains images with poor lighting conditions. Which factors should you investigate to identify relationships or trends that may affect the model's performance?
- A
The distribution of lighting conditions in the training dataset compared to the validation dataset.
- B
The GPU utilization percentage during training and inference tasks.
- C
The augmentation techniques applied to the training data, especially for lighting variations.
- D
The choice of NVIDIA GPU architecture used for training the model.
- E
The hyperparameters used for model optimization, such as learning rate and batch size.
Show answer and explanation
Correct answers: A, C, E
Explanation
To identify relationships and trends affecting the model's performance, it's essential to examine the data characteristics (e.g., lighting conditions in training vs. validation datasets), the preprocessing and augmentation techniques, and the hyperparameters used during training. These factors directly influence the model's ability to generalize to unseen data. GPU architecture and utilization, while important for performance efficiency, are not directly relevant to the model's ability to handle specific data conditions like poor lighting.
- A. Correct.
Investigating the distribution of lighting conditions in the training dataset compared to the validation dataset is essential. If the training dataset does not include enough examples with poor lighting, the model will struggle to generalize to such scenarios.
- B. Incorrect.
GPU utilization percentage during training and inference tasks is unrelated to the model's ability to handle variations in lighting conditions. This factor primarily impacts performance efficiency rather than model accuracy.
- C. Correct.
Examining the augmentation techniques is critical, as appropriate augmentations (e.g., brightness adjustments) can help the model generalize to a variety of lighting conditions.
- D. Incorrect.
While the choice of NVIDIA GPU architecture can affect training speed and resource utilization, it does not directly influence the model's ability to handle lighting variations in the data.
- E. Correct.
Hyperparameters such as learning rate and batch size can significantly influence the model's convergence and generalization. Poorly tuned hyperparameters might prevent the model from learning robust features, including those required to handle variations in lighting.