Meta Learning v.s Machine learning-CSDN博客

本文链接：https://blog.csdn.net/timingzj/article/details/129333392

Machine Learning = Looking for a function

Function with unknown $\small f_{\theta }$ $\small \theta$ :the unknown parameters
Define loss function $\small L(\theta) = \sum_{k=1}^{K} e_{k}$ ( $\small L(\theta )$ :sum over training examples in one task)
Optimization $\small \theta _{}^{*} = arg\, \min_{\theta}^{}L(\theta )$ $\small f_{\theta _{}^{*}}$

What is learnable in a learning algorithm? Net Architecture, Initial Parameters, Learning Rate, ... $\phi$ :learnable components
Define loss function ( $\small L(\phi )$ ) for learning algorithm $\small F_{\phi }$ $\small L(\phi ) = \sum_{n=1}^{N} l_{}^{n}$ the smaller the better, N is the number of the training tasks ( $\small l_{}^{n}$ :sum over testing examples in one task; $\small L(\phi )$ :sum over training tasks)
Find $\small \phi$ that can minimize $\small L(\phi )$ $\small \phi _{}^{*} = arg\,\min_{\phi }^{}L(\phi )$ , then have a learned "learning algorithm" $\small F_{\phi_ {}^{*}}$