On sensitivity of meta-learning to support data
On sensitivity of meta-learning to support data
复制标题
DOI：
--
复制
发表时间：
2021-10
期刊：
影响因子：
--
通讯作者：
Mayank Agarwal;M. Yurochkin;Yuekai Sun
中科院分区：
文献类型：
--
作者：
Mayank Agarwal;M. Yurochkin;Yuekai Sun
元学习算法被广泛用于少量学习。例如，图像识别系统在只看到几个标记的例子后就很容易适应看不见的类。尽管取得了成功，但我们表明现代元学习算法对用于适应的数据（即支持数据）极其敏感。特别是，我们证明了存在（未改变的，在分布，自然）的图像，当用于适应，产量精度低至4\%或高达95\%的标准少拍图像分类基准。我们解释了我们的实证研究结果的类利润率，这反过来表明，强大和安全的元学习需要更大的利润比监督学习。
Meta-learning algorithms are widely used for few-shot learning. For example, image recognition systems that readily adapt to unseen classes after seeing only a few labeled examples. Despite their success, we show that modern meta-learning algorithms are extremely sensitive to the data used for adaptation, i.e. support data. In particular, we demonstrate the existence of (unaltered, in-distribution, natural) images that, when used for adaptation, yield accuracy as low as 4\% or as high as 95\% on standard few-shot image classification benchmarks. We explain our empirical findings in terms of class margins, which in turn suggests that robust and safe meta-learning requires larger margins than supervised learning.