Less training data, Less model performance. Is it inevitable that pre-training + few-shot learning will not be as good as sufficient data in a specific field?
It really depends on the content of the problem and the objective of solving it. In some cases, what we want is to reduce the training time, reduce computational complexity, and increase the convergence rate so that the trained model can be reused again without fine-tuning the whole network. As long as the performance hits the minimum required performance, it is good. However, the amount of data, the correlation between different tasks and their corresponding data, and the number of tasks really matter. If you train a model that is going to be reused for several different tasks with different data, of course there will be some performance degradation due to the effect of generalization. However, if there are only two tasks (or two environments) that are very similar and you only want to reduce the training time by applying different transfer learning or meta learning methods, it is possible to reach the same performance with the model that is fully fine-tuned.
The main question is what is the problem statement. If we want the machine to understand, it is better to train the model with large data, only then we can rely on its results.
Transfer Learning will be suited to solve this problem I think you should try it
the process of data argumentation seeks to vary the nature of already existing train data so as to help increase the effectiveness of your model or algorithm
- Badges
- Science topic
- Similar topics
- Computer Science and Engineering
- Programming Languages
- Python