On sensitivity of meta-learning to support data
This reveals a critical vulnerability in meta-learning for few-shot image classification, which is incremental as it builds on existing methods to expose a specific issue.
The paper tackles the problem of meta-learning algorithms being highly sensitive to support data in few-shot learning, showing that using different natural images for adaptation can cause accuracy to vary from 4% to 95% on standard benchmarks.
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.