Learning to Learn by Modular Neural Networks

1
Learning to Learn by Modular Neural Networks
Mojuurugata nyuurarunetto ni yoru gakusyuu no
tame no gakusyuu

• Akio Sashima, Kazuo Hiraki
• Electrotechnical Laboratory
• sashima,khiraki_at_etl.go.jp

2
Abstract

• Encountering a stream of learning tasks, humans
  learn not only knowledge of current task but also
  biases of learning future tasks. To model this
  ability, we focus attention on modular
  representation of cognitive system and
  hypothesize that the expert modules become a key
  bias for learning new task. We examine this
  hypothesis by computer simulation using
  artificial modular neural networks.

3
Learning to Learn

• To learn complex tasks, learn simple tasks first.
  (e.g. motor skill, mathematics).
• Encountering a stream of learning tasks,
• Most machine learning algorithms only learn
  knowledge of current task.
• Humans learn not only knowledge of current task
  but also biases for learning new tasks.
• ?learning to learn
• What kind of cognitive system shows ability in
  learning to learn?

4
Modular Representation of Cognitive System

• Modularity of MindFodor 83.
• Human cognitive system consists of innate
  functional modules and a central system.
• Representational Rediscription Karmiloff-Smith
  92.
• Developmental stages of children correspond to
  modular representation of their cognitive system.
• Modularity is developmentally acquired.

5
Learning to Learn with Modular Representation

• We assume that
• Expert modules are acquired with learning.
• Expert modules for simple function are acquired
  with learning simple tasks.
• Expert modules for simple function are reusable.
• Using reusable expert modules, cognitive system
  learns complex task easily.

6
Modular Neural Networks

• Modular neural networks consist of expert modules
  and gating moduleJacobs 91Wolpert 98.

7
Experiments Function Approximation Task
8
Learning Two Functions Continuously

• AB Task
• The networks learn function A first, and then
  learn function B.
• BB Task
• The networks learn function B in twice.

9
Network Architecture
10
Experimental Results

• By all trials(20 times), the networks at AB task
  could approximate function B, but the networks at
  BB task could not approximate function B.
• Expert modules, which were acquired with learning
  function A, make learning function B easier.
• Modular representation could be a key factor of
  ability in learning to learn.

11
Future Works

• Experiments with biological plausible tasks.
• (e.g. arm control,eye movement.)
• Modeling developmental process as learning to
  learn.
• (e.g. On each developmental stage, humans may
  learn biases on next stage.)