Prototypical Networks for Few-shot Learning. NIPS 2017

• Prototypical Networks for Few-shot Learning. NIPS 2017
  • Abstract
  • Prototypical Networks
    • Model

Abstract

• some simple design decisions can yield substantial improvements over recent approaches involving complicated architectural choices and meta-learning.
• Prototypical Networks learn a metric space in which classification can be performed by computing distances to prototype representations of each class.

Prototypical Networks

Model

• Prototype \(c_k\) for class \(k\): \(c_k = \frac{1}{|S_k|}\sum_\limits{(x_i, y_i)\in S_k} f_\phi(x_i) \tag 1\)
• Predicted probability for class \(k\): \(p_\phi(y=k|x) = \dfrac{\exp(-d(f_\phi(x), c_k))}{\sum_{k'}\exp(-d(f_\phi(x), c_{k'}))} \tag 2\)
• Loss of one query data: \(\dfrac{1}{N_CN_Q}[d(f_\phi(x),c_k)+\log\sum_{k'}\exp(-d(f_\phi(x),c_{k'}))]\tag 3\)

• \(D\) 是整個 trainin set；\(D_k\) 是整個 training set 中的 class \(k\) data
• \(f_\phi\) 是 embedding function
• Update loss 那項 本來是 \(-\log(\exp(-d(f_\phi(x), c_k)))\)，化簡後得到 \(d(f_\phi(x), c_k)\)

• support set 裡面同 class 的 image 的 embedding 的 mean 被稱作 prototype
• 訓練使得 query set 到自己類別 prototype 的距離越近越好；到其他類別 prototype 距離越遠越好
• 有做 additional 實驗，training task 用的 way 比 testing task 更多，效果比較好；而使用的 shot 和 testing task 一樣，效果比較好。
• 使用 Euclidean distance 作為距離，表現比 cosine similarity 好，推測是因為 cosine similarity 不是 Bregman divergence (???)
• 也可做 zero-shot learning

tags: fewshot learning