When a suitable initial parameter set $$\theta_0$$ is not available, but parts of $$\theta_0$$ are accessible from related pre-trained models, these parts can be aggregated to form a suitable $$\theta_0$$. This aggregated parameter set can then be used directly or fine-tuned using the training dataset $$D_{train}$$.

University of Notre Dame

Google

When using $D_{train}$ to refine $\theta_0$, there’re several techniques can be used to improve results:

1)	Fine-Tuning Existing Parameter by Regularization

2)	Aggregating a Set of Parameters.

3)	Fine-Tuning Existing Parameter with New Parameters.


Strategy of Fine Tuning Parameters 

Wang, Y., Yao, Q., Kwok, J. T., & Ni, L. M. (2020). Generalizing from a few examples: A survey on few-shot learning. ACM computing surveys (csur), 53(3), 1-34.

Generalizing from a Few Examples: A Survey on Few-Shot Learning

Given the few-shot dataset and model, just tuning $\theta_0$ using gradient descent can easily lead to overfitting. Thus some regularization techniques are used, including early-stopping, selectively updating, updating related parts of $\theta_0$ together, and model regression network.

Fine-Tuning Existing Parameter by Regularization

When $\theta_0$ alone is not enough to encode the input of the new few-shot learning task, we may add a new parameter to the current $\theta_0$, and fine-tune $\theta_0$ while learning new parameters.

Learn Before

Related