DWQA QuestionsCategory: DatabaseFor participants with different amounts of data, does it make sense to compare their loss values under the same training network, or how to compare their loss values?
zack asked 1 month ago

The specific problem situation is like this
The network structure is based on the simplest assumption of multi-layer perceptron.
a. B and C have different data volumes 1E2, 1E4 and 1e6. They have the same network structure model (the model is exactly the same), and they use the same loss_ Function: three people train their own models, and there is a loss value in the training process.
Due to the different amount of data, the value of loss will be different, and the data volume is relatively large.
If it is only a simple $/ frac {loss} {data volume} $, the party with large data volume will be too small no matter how much the loss itself is due to the small range of change in the value of loss
How to compare the loss values of the three? (if it is meaningless, why not

1 Answers
zack answered 1 month ago

Personal feelings,
The meaning of loss is to indicate whether the training is normal, while test accuracy means whether the model performs well on the task.
Therefore, it is meaningless to compare loss according to this point.