06.09.2021 Views

Mind, Body, World- Foundations of Cognitive Science, 2013a

Mind, Body, World- Foundations of Cognitive Science, 2013a

Mind, Body, World- Foundations of Cognitive Science, 2013a

SHOW MORE
SHOW LESS

You also want an ePaper? Increase the reach of your titles

YUMPU automatically turns print PDFs into web optimized ePapers that Google loves.

input signals coming from all <strong>of</strong> the units. The connection weights used by the<br />

hidden units place meaningful constraints on how these signals are combined.<br />

Let us consider the role <strong>of</strong> the particular connection weights used by the hidden<br />

units. Given the binary nature <strong>of</strong> the input encoding, the net input <strong>of</strong> any hidden<br />

unit is simply the sum <strong>of</strong> the weights associated with each <strong>of</strong> the activated input<br />

units. For a value unit, if the net input is equal to the value <strong>of</strong> the unit’s m then the<br />

output generates a maximum value <strong>of</strong> 1.00. As the net input moves away from m in<br />

either a positive or negative direction, activity quickly decreases. At the end <strong>of</strong> training,<br />

the values <strong>of</strong> m for the three hidden units were 0.00, 0.00, and –0.03 for Hidden<br />

Units 1, 2, and 3, respectively. Thus for each hidden unit, if the incoming signals are<br />

essentially zero—that is if all the incoming signals cancel each other out—then high<br />

activity will be produced.<br />

Why then do Hidden Units 1 and 2 use the same set <strong>of</strong> four connection weights<br />

but assign these weights to different sets <strong>of</strong> input notes? The answer is that these<br />

hidden units capture similar chord relationships but do so using notes from different<br />

strange circles.<br />

This is shown by examining the responses <strong>of</strong> each hidden unit to each input<br />

chord after training. Table 4-3 summarizes these responses, and shows that each<br />

hidden unit generated identical responses to different subsets <strong>of</strong> input chords.<br />

Input Chord<br />

Activation<br />

Chord Chord Root Hid1 Hid2 Hid3<br />

Major<br />

C, D, A, F#, G#, A# 0.16 0.06 0.16<br />

C#, D#, F, G, A, B 0.06 0.16 0.16<br />

Major7<br />

C, D, A, F#, G#, A# 0.01 0.12 1.00<br />

C#, D#, F, G, A, B 0.12 0.01 1.00<br />

Dom7<br />

C, D, A, F#, G#, A# 0.27 0.59 0.00<br />

C#, D#, F, G, A, B 0.59 0.27 0.00<br />

6th<br />

C, D, A, F#, G#, A# 0.84 0.03 1.00<br />

C#, D#, F, G, A, B 0.03 0.84 1.00<br />

Table 4-3. The activations produced in each hidden unit by different subsets <strong>of</strong><br />

input chords.<br />

From Table 4-3, one can see that the activity <strong>of</strong> Hidden Unit 3 is simplest to describe:<br />

when presented with a dominant 7th chord, it produces an activation <strong>of</strong> 0 and a weak<br />

activation to a major triad. When presented with either a major 7th or a 6th chord, it<br />

produces maximum activity. This pattern <strong>of</strong> activation is easily explained by considering<br />

the weights that feed into Hidden Unit 3 (Table 4-2). Any major 7th or 6th chord<br />

is created out <strong>of</strong> two notes from one circle <strong>of</strong> major seconds and two notes from the<br />

Elements <strong>of</strong> Connectionist <strong>Cognitive</strong> <strong>Science</strong> 169

Hooray! Your file is uploaded and ready to be published.

Saved successfully!

Ooh no, something went wrong!