The neural marketplace
File(s)
Author(s)
Lewis, Sarah Noami
Type
Thesis
Abstract
The `retroaxonal hypothesis' (Harris, 2008) posits a role for slow retrograde signalling in learning. It is based on the intuition that cells with strong output synapses tend to be those that encode useful information; and that cells which encode useful information should not modify their input synapses too readily. The hypothesis has
two parts: rst, that the stronger a cell's output synapses, the less likely it is to change its input synapses; and second, that a cell is more likely to revert changes to its input synapses when the changes are followed by weakening of its output synapses. It is motivated in part by analogy between a neural network and a market economy, viewing neurons as `entrepreneurs' who `sell' spike trains to each other. In this view, the slow retrograde signals which tell a neuron
that it has strong output synapses are `money' and imply that what it produces is useful.
This thesis constructs a mathematical model of learning, which validates the intuition of the retroaxonal hypothesis. In this model, we show that neurons can estimate their usefulness, or `worth', from the magnitude of their output weights. We also show that by making each cell's input synapses more or less plastic according to its worth, the performance of a network can be improved.
two parts: rst, that the stronger a cell's output synapses, the less likely it is to change its input synapses; and second, that a cell is more likely to revert changes to its input synapses when the changes are followed by weakening of its output synapses. It is motivated in part by analogy between a neural network and a market economy, viewing neurons as `entrepreneurs' who `sell' spike trains to each other. In this view, the slow retrograde signals which tell a neuron
that it has strong output synapses are `money' and imply that what it produces is useful.
This thesis constructs a mathematical model of learning, which validates the intuition of the retroaxonal hypothesis. In this model, we show that neurons can estimate their usefulness, or `worth', from the magnitude of their output weights. We also show that by making each cell's input synapses more or less plastic according to its worth, the performance of a network can be improved.
Version
Open Access
Date Issued
2016-06
Date Awarded
2016-12
Copyright Statement
Attribution NoDerivatives 4.0 International Licence (CC BY-ND)
Advisor
Harris, Kenneth
Clopath, Claudia
Publisher Department
Bioengineering
Publisher Institution
Imperial College London
Qualification Level
Doctoral
Qualification Name
Doctor of Philosophy (PhD)