Combined Model and Task Learning, and Other Mechanisms - Computational Explorations in Cognitive Neuroscience - page 181

Hidden

Christo=Penny Andy=Christi

Marge=Art Vicky=James Jenn=Chuck

Colin Charlot

Rob=Maria Pierro=Francy

Gina=Emilio Angela=Tomaso

Lucia=Marco

Alf Sophia

Figure 6.7: The family tree structure learned in the family
trees task. There are two isomorphic families, one English
and one Italian. The = symbol indicates marriage.

Agent_Code Relation_Code Patient_Code

6.4.1 Exploration of a Deep Network Relation

Agent

Patient

Now, let's explore the case of learning in a deep network
using the same family trees task as O'Reilly (1996b)
and Hinton (1986). The structure of the environment is
shown in figure 6.7. The network is trained to produce
the correct name in response to questions like “Rob is
married to whom?” These questions are presented by
activating one of 24 name units in an agent input layer
(e.g., “Rob”), in conjunction with one of 12 units in a
relation input layer (e.g., “Married”), and training the
network to produce the correct unit activation over the
patient output layer.

Figure 6.8: The family tree network, with intermediate Code
hidden layers that re-represent the input/output patterns.

This will bring up a window displaying the first ten
training events, which should help you to understand
how the task is presented to the network.

Go ahead and scroll down the events list, and see
the names of the different events presented to the net-
work (you can click on any that look particularly interest-
ing to see how they are represented).
Now, let's see how this works with the network itself.

, !

Open project family_trees.proj.gz in

chapter_6 .
First, notice that the network (figure 6.8) has Agent
and Relation input layers, and a Patient output
layer all at the bottom of the network. These layers
have localist representations of the 24 different peo-
ple and 12 different relationships, which means that
there is no “overt” similarity in these input patterns
between any of the people. Thus, the Agent_Code ,
Relation_Code ,and Patient_Code hidden lay-
ers provide a means for the network to re-represent
these localist representations as richer distributed pat-
terns that should facilitate the learning of the mapping
by emphasizing relevant distinctions and deemphasiz-
ing irrelevant ones. The central Hidden layer is re-
sponsible for performing the mapping between these re-
coded representations to produce the correct answers.

, !

Press the Step button in the control panel.
The activations in the network display reflect the mi-
nus phase state for the first training event (selected at
random from the list of all training events).

, !

Press Step again to see the plus phase activations.
The default network is using a combination of
Hebbian and GeneRec error-driven learning, with the
amount of Hebbian learning set to .01 as reflected by
the lrn.hebb parameter in the control panel. Let's
see how long it takes this network to learn the task.

, !

Open up a graph log to monitor training by press-
ing View , TRAIN_GRAPH_LOG , turn the Display of the
network off, and press Run to allow the network to train
on all the events rapidly.
As the network trains, the graph log displays the er-
ror count statistic for training (in red) and the average
number of network settling cycles (in orange).

, !

Press View and select EVENTS on the

family_trees_ctrl control panel.

, !