Combined Model and Task Learning, and Other Mechanisms - Computational Explorations in Cognitive Neuroscience - page 171

Chapter 6

Combined Model and Task Learning, and
Other Mechanisms

6.7.2 Phase-Based Temporal Differences . . . . . 198
6.7.3 Exploration of TD: Classical Conditioning . 199
6.8 Summary ......................202
6.9 FurtherReading ..................2 02

Contents

6.1 Overview ......................173
6.2 Combined Hebbian and Error-driven Learning . 173

6.2.1 Pros and Cons of Hebbian and Error-
DrivenLearning...............174
6.2.2 Advantages to Combining Hebbian and
Error-DrivenLearning ...........175
6.2.3 Inhibitory Competition as a Model-
LearningConstraint.............175
6.2.4 Implementation of Combined Model and
TaskLearning................176
6.2.5 Summary ..................177
6.3 GeneralizationinBidirectionalNetworks.....178
6.3.1

6.1 Overview

Having developed and explored two different kinds of
learning in the previous two chapters (Hebbian model
learning in chapter 4 and error-driven task learning in
chapter 5), we now explore their combination into one
unified learning model. This model provides the ad-
vantages of both forms of learning, and it demonstrates
some important synergies that arise from their combi-
nation.
We also explore two other kinds of learning mech-
anisms that are necessary to deal with tasks that: (1)
involve temporally extended sequential processing, and
(2) have temporally delayed contingencies. Both of
these mechanisms can be implemented with simple ex-
tensions of our basic model and task learning frame-
work, and both have a clear biological basis.

ExplorationofGeneralization........179

6.4 Learning to Re-represent in Deep Networks . . . 181

6.4.1 ExplorationofaDeepNetwork.......183

6.5 Sequence and Temporally Delayed Learning . . . 186

6.6 Context Representations and Sequential Learning 187

6.6.1

Computational Considerations for Context
Representations ...............188

6.6.2 Possible Biological Bases for Context Rep-

resentations .................189

6.2 Combined Hebbian and Error-driven

Learning

6.6.3 Exploration: Learning the Reber Grammar . 189

6.6.4 Summary ..................193

6.7 Reinforcement Learning for Temporally De-

layedOutcomes...................193
6.7.1

In this section, we address two important and related
questions: (1) What are the relative advantages and dis-
advantages of Hebbian and error-driven learning, and

TheTemporalDifferencesAlgorithm ....195

173