If we train a model every time from scratch by using current task and samples from memory (ER) then is it correct way to perform continual learning?

Suppose that there are T tasks. We use an experience replay (ER) strategy using a tiny episodic memory. Here, we train a model always from scratch at each task using current task samples and samples from memory. However, this model works perfectly fine for previous and current tasks.

Whether this way of performing continual learning is correct or not as we are not training the previous model $(t^{th})$ continually for the next task $((t+1)^{th})$? Are we violating the continual learning norms?

Topic online-learning neural-network machine-learning

Category Data Science

1
answered at

Although both methods are correct, it's better to train the old model since we already trained one model, and starting over from scratch would cost more. The old model already had optimal weights, so with new data, it only needs to fine-tune the model.

The model performance is not affected by these approaches as long as we have the same distributed train data. However, we optimize computation costs as data scientists or ML engineers. ImageNet is an amazing model that gives good results when used with already trained data. However, if we want to use different data, we must train the model but only need to fine-tune it. From scratch requires a lot of time and money.

Conclusion: Model performance will not be affected by these approaches, but training from scratch requires more time and resources.

About

Geeks Mental is a community that publishes articles and tutorials about Web, Android, Data Science, new techniques and Linux security.