3个组成
It employs a learner, two training phases, and one testing phase.
1. learner
The learner is assumed to be composed of a feature extractor and a multi-class classifier.
2. traing phased one (representation learning)
During representation learning (training phase one), the
learner receives a fixed set of base categories C b a s e C_{base} Cbase, and a dataset D containing a large number of examples for each category in Cbase. The learner uses D to set the parameters of its feature extractor.
产生新的类别
Any two examples z1 and z2 belonging to the same category represent a plausible transformation. Then, given a novel category example x, we want to apply to x the transformation that sent z1 to z2. That is, we want to complete the transformation “analogy” z1 : z2 :: x : ?.
We do this by training a function G that takes as input the concatenated feature vectors of the three examples [φ(x), φ(z1), φ(z2)].G是MLP
3. second phase (low-shot learning)
the learner is given a set of categories C l C_l Cl that it must learn to distinguish. C l C_l Cl = C b a s e C_{base} Cbase ∪ \cup ∪ C n o v e l C_{novel} Cnovel is a mix of base categories Cbase, and unseen novel categories Cnovel.
对于 C n o v e l C_{novel} Cnovel,仅仅有k-shot
4.testing phase
the learnt model predicts labels from the combined label space C l C_l Cl = C b a s e C_{base} Cbase ∪ \cup ∪ C n o v e l C_{novel} Cnovel on a set of previously unseen test images.
模型(Learning to generate new examples)
核心:,通过z1:z2,类比x:?.
利用D训练G, 为用于类比的2组双胞胎。预测结果为:
。损失函数
。与上文中的特征表达误差与分类误差对应。