NCA-AIIO Question 87
Select 3An AI research team is training a deep learning model for image recognition using a dataset of human facial images. During the training process, the team notices that the model's accuracy on the validation set is significantly lower than on the training set. Which factors could be contributing to this issue, and how should they be addressed?
- A
The dataset contains class imbalances, where certain facial features or demographics are underrepresented.
- B
The training set is much larger than the validation set, leading to overfitting.
- C
The model's architecture is too simple and lacks the capacity to learn complex patterns in the data.
- D
The dataset includes mislabeled samples in the training set, introducing noise into the model.
- E
The team is using a learning rate that is too high, causing the model to fail to converge.
Show answer and explanation
Correct answers: A, B, D
Explanation
The lower validation accuracy compared to the training accuracy suggests issues such as overfitting, data imbalance, or noise in the training data. Addressing these specific factors is critical to improving the model's generalization ability and ensuring robust performance on unseen data.
- A. Correct.
Class imbalances can cause the model to focus more on overrepresented classes, leading to bias and poor generalization performance on the validation set. This would need to be addressed through techniques such as data augmentation or re-sampling.
- B. Correct.
If the training set is much larger than the validation set and the model performs well on the training data but poorly on validation, it is an indication of overfitting. A larger validation set or regularization techniques can help mitigate this.
- C. Incorrect.
An overly simplistic model architecture would likely underfit the training data, resulting in poor performance on both the training and validation sets, which is not the issue described here.
- D. Correct.
Mislabeled samples in the training set introduce noise and can negatively impact the model's ability to generalize to the validation set. Cleaning the dataset or relying on noise-robust training techniques can help mitigate this.
- E. Incorrect.
A high learning rate typically leads to divergence or instability during training, which is not consistent with the described scenario of high training accuracy but low validation accuracy.