From the variation of validation loss, no overfitting occurs during the training process.
When you increase the data set, do the training set and validation set increase? Or, only the training set. It looks like overfitting in this case, the model mainly focus on the training set, and I think it might be the optimal point for your research. For example, when we want to complete various information, like A, B, and C, if you only provide A, the performance must be lower than the expected performance. Hoping it is useful for you!
Jiarong Chen Thanks for your answer！Only the training set increases in this case. Two sets of experiments were done, one(E1) containing a sampled training set and the other(E2) containing an unsampled training set. The validation set remains the same in these two experiments. The result shows E2 has a lower training loss, but the validation loss is almost the same as that in E1.
Get it！It seems that E1+E2 is not better than E1, and E2 might be useless. I think you can try E2 for training, and discuss the different results. Do these datasets exist the domain gap?
Jiarong Chen Thanks again! I will try your advice. And there is no domain gap between these datasets. I guess that the data distribution of E2 is similar to that of E1, increasing the training set does not improve the model performance, though the training loss has a lower value.
- Badges
- Science topic