Bayesian Meta-Learning

What does Theta Mean

Training and testing must match. Task must share structure.

The “Structure” means statistical dependence on shared latent information $\theta$.

Bayesian Meta-Learning Approaches

第一个想到的方法是让模型直接输出关于$y\^{ts}$分布的参数值。

好处是简单、能够结合其他的多种方法。

坏处是不能得到模型函数的不确定性原因,如确定数据点之间的不确定性如何关联。只能够表达有限的相对于目标$y^{ts}$的分布类别。倾向于产生低校准的不确定性估计。

The Bayesian Deep Learning Toolbox

The goal is to represent distributions with neural networks.

Bayesian black-box meta-learning

Bayesian optimization-based meta-learning

Use ensembles to model non-gaussian posterior

Sample Parameter Vectors

Sample parameter vectors with a procedure like Hamiltonian Monte Carlo to model non-Gaussian posterior over all parameters:

Methods Summary

How to Evaluate a Bayesian Meta-Learner

Use the standard benchmarks i.e. MiniImagenet accuracy:

Pros:

  • Standardized
  • Real Images
  • Good check that the approach didn’t break anything

Cons:

  • Metrics like accuracy don’t evaluate uncertainty
  • Tasks may not exhibit ambiguity
  • Uncertainty may not be useful on this dataset

Note: Cover Picture