The video classification is based on a combination of CNN and RNN, and the correct answer rate is low in the test data, while the training data shows an almost 100% correct answer rate.
However, the correct answer rate of the test data did not drop steadily in the middle like the typical graph of overlearning.
Currently, we have about 50 data, which is quite small and has a relatively large distribution between them.
Can I think of it as overlearning when the correct rate of test data is low for training data even if there is a small number of data like this?
Or should we think that there is another cause?
I would appreciate it if you could let me know
I think the following confirmation is mainly necessary.
How many of the 1.50 data are used for testing?
2. No matter how many times you try, will the result be the same every time?
3. Is the method of separating training and test data the same every time?
4. Is the amount of data sufficient according to the number of output classifications?
5. What will happen if both the training and the test are combined to ensure that the correct answer rate will be higher?
6. What will happen if the number of training sessions is reduced?
In your question, "Determining if overlearning is occurring when the number of data is small," it is important to check 6 of the above 1-6, but the cause cannot be determined by itself.
Each of them explains
1. Extremely high or low test data, for example, is not a proper training or test.
2. In some cases, the test data may be too small.
3. If you divide it the same way every time, it may be that the test is fixed with a combination of patterns with a low correct answer rate.
4. Too little data will not provide sufficient versatility.This can be said to be overlearning, but it is rather a mere lack of training variations.As the number of classifications increases, the amount of data required may increase suddenly, so you need to be careful.In this case, I think the amount of data is probably not enough.
5. If the results do not change after doing this, there is a possibility that there is a bug in the program itself or that what you are trying to do is unreasonable.
6. If versatility increases, it can be overlearning.
If 1-5 is fine and 6 is more versatile, then there is a high possibility of overlearning.
Looking at the questioner's situation, it is true that the generalization capability of the model has been lost and is over-learning.
However, since the number of data is about 50, depending on how the test data is divided, the training data may be biased.
© 2024 OneMinuteCode. All rights reserved.