Differentiable plasticity: training plastic neural networks withbackpropagation

http://proceedings.mlr.press/v80/miconi18a/miconi18a.pdf

Abstract

1、 通过梯度下降 优化可微塑性

2、 在测试集合中， 训练集中从未见过的自然图像集合， 能重建。

3、 可以解决一般的元学习任务。

Introduction

Many of the recent spectacular successes in machine learn-ing involve learning one complex task very well, throughextensive training over thousands or millions of trainingexamples (Krizhevsky et al., 2012; Mnih et al., 2015; Sil-ver et al., 2016).  After learning is complete, the agent’sknowledge is fixed and unchanging; if the agent is to beapplied to a different task, it must be re-trained (fully or par-tially), again requiring a very large number of new training examples. By contrast, biological agents exhibit a remark-able ability to learn quickly and efficiently from ongoingexperience: animals can learn to navigate and remember thelocation of (and quickest way to) food sources, discover andremember rewarding or aversive properties of novel objectsand situations, etc. – often from a single exposure.

Endowing artificial agents with lifelong learning abilitiesis essential to allowing them to master environments withchanging or unpredictable features, or specific features thatare unknowable at the time of training. For example, super-vised learning in deep neural networks can allow a neuralnetwork to identify letters from<