Prototypical Networks for Few-shot Learning

 

This text is study for paper papers.nips.cc/paper/2017/file/cb8da6767461f2812ae4290eac7cbc42-Paper.pdf

Overview

• Propose Prototypical netwroks for the problem of few-shot classification, where a classifier must generalize to not seen class in the training set, given only a small number of examples of each new classes
• Model learns a metric space by computing distances to protytype respresentations of each class

Prototypical Networks

• Notation
  • Support set of N labeled examples S = { (x_1,y_1) ,..., (x_N,y_N) }
  • x_i ∈ R^D, which is D-dimensional feature vector
  • y_i ∈ {1,...,K} is the correspoinding label
  • S_k is the set of examples labeled with class k
• Model
  • Prototypical networks compute prototype c_k∈ R^M, through embedding function f_ϕ
  • f_ϕ : R^D -> R^M with learnable parameter ϕ
  • Each prototype is the mean vector of the embedded support points belonging to its class
  • Given a distance function d d : R^M *R^M ->[0 ,+∞), prototypical networks produce a distribution over classes for a query point x based on softmax distances 
  • Learning proceeds by minimizing the negative log-probability J(ϕ)=-logp_ϕ(y=k|x)