Lightweight fine-tuning is one of the most important techniques for adapting foundation models, because it allows you to modify foundation models for your needs without needing substantial computational resources.
0
stars
0
forks
source link
Feedback: training is not required for the foundational model #2
My approach of training a pretrained model would be suitable e.g. for a conventional model like a CNN in a classification task, transfer learning, etc., however foundational models have exponentially more parameters, and those parameters are trained on millions of datapoints, some on all available text in the Internet. Retraining those parameters, because you have e.g. 1000 labeled sentences in a training dataset is not reasonable and you run into risk of as mentioned catastrophic forgetting.
My approach of training a pretrained model would be suitable e.g. for a conventional model like a CNN in a classification task, transfer learning, etc., however foundational models have exponentially more parameters, and those parameters are trained on millions of datapoints, some on all available text in the Internet. Retraining those parameters, because you have e.g. 1000 labeled sentences in a training dataset is not reasonable and you run into risk of as mentioned catastrophic forgetting.