Behavior-Based Early Language Development on a Humanoid Robot



in the vocal label is high, Concept transitions to the
UpdateGains state.

If the activation level reaches above a threshold,
the behavior transitions to the
OutputLabel state,
which sets the behavior’s priority to a value de-
pendent on its activation and the confidence of the
match, sets the output priority to the behavior’s pri-
ority, and outputs an inhibitory signal at that pri-
ority level. The inhibitory outputs of each concept
connect back to the
HeardThis behavior. Such a
signal received by
HeardThis propagates further to
ConceptMap and inhibits the default creation of a new
Concept when a new string of phonemes is heard (see
section 5.2).

5.1.2 Updating of label and confidence

Whether or not the vocal label of a Concept should
be updated is determined by a combination of con-
ditions on the
Concept’s confidence C, the confidence
Cheard in the accuracy of the heard phonemic string,
the value of the best match between them and the
current activation
A of the Concept. The update la-
bel value
UL is compared against a threshold to de-
cide whether an update should take place (where
k is
a scaling constant, chosen arbitrarily, and
θ the acti-
vation threshold,
d is the distance measure between
the two strings in the result of the
B estM atch() call):

C C C heard Л

Ul = ʌ k c 0

A>θ, C 6=0
otherwise


(3)


If the Concept is active, then UL will be proportional
to the confidence attached to the heard string, and
inversely proportionate to the goodness of the match
and to the level of the
Concept’s own confidence. If it
is not active, there is no reason to update the label.

These computations depend on the confidence
measure on the input speech string. This is com-
puted inside
HeardThis in the following way:

Cheard = default - n X d~+ι Cn    (4)

n

where def ault is a constant, dn is a distance mea-
sure of the match between
heard and the n’th tem-
plate,
Cn is the confidence for the n’th template,
and the sum is taken over all templates that had at
least a minimal match to the heard string. The in-
dividual
Cn and dn are transmitted to the HeardThis
behavior through incoming connections. N of these
connections are created and we then let the active
Concept behaviors compete to overwrite those con-
nections only.

Note that Cn are used to compute Cheard but
Cheard is needed to determine the update rule for
the vocal label of concept
n. However, this situation
does not start an infinite regression in our implemen-
tation, as there is a clear time difference between
the activation of the two connections, so that when
Cheard is received back at Concept, the behavior has
switched to another state as shown in figure 6.

5.1.3 Updating of parameters

When a behavior is created, the gains are set to
default values, which must be updated in state
UpdateGains to arrive at some consistent represen-
tation of releaser properties for that behavior. The
signals that the
Concept has access to at this point
in the state machine include:

its activation A

the result of the matching process between its
vocal label and the input phonemic string

its confidence in its vocal label C

a measure of confidence in the accuracy of the
string heard
Cheard

the current presence or absence of relevant re-
leasers

From this information the behavior derives a rein-
forcement signal
UG computed as follows, where d
is the distance measure returned by BestMatch(),
which is equal to 0 when the match is perfect:

- k Cheard Cd

UG = S        Cheard C

I             d + 1

A>θ
otherwise


(5)


This rule allows for both Type I and Type II errors:
when the behavior is active even though something
other than its label is heard, and when the behavior
remains inactive and its label was heard. However,
in the first case reinforcement should be weaker since
it is quite often the case that the speech the robot
hears is not describing any immediate features of the
environment. Therefore we scale this type of signal
by a smaller
k. This does not eliminate the problem,
but biases the computation in a simple way.
UG is
then used in the update rule for each of the gains in
the vector:

Git+1 = Git + αRit Git UG           (6)

This will not change the value of Gi (the ith element
of the gains vector
G) if its corresponding Receptor
Ri was not responding.

HeardThis

J⅞eard


t+1


I inhibition

Concept

t+2

t+3


t+4


t+5


Figure 6: The timings of information transfer between
HeardThis and any Concept. The state transitions ensure
that confidence measures are only computed once.



More intriguing information

1. Input-Output Analysis, Linear Programming and Modified Multipliers
2. Clinical Teaching and OSCE in Pediatrics
3. The name is absent
4. AGRICULTURAL PRODUCERS' WILLINGNESS TO PAY FOR REAL-TIME MESOSCALE WEATHER INFORMATION
5. Reform of the EU Sugar Regime: Impacts on Sugar Production in Ireland
6. Barriers and Limitations in the Development of Industrial Innovation in the Region
7. Strategic Effects and Incentives in Multi-issue Bargaining Games
8. Consumer Networks and Firm Reputation: A First Experimental Investigation
9. The name is absent
10. Education Responses to Climate Change and Quality: Two Parts of the Same Agenda?
11. The name is absent
12. Uncertain Productivity Growth and the Choice between FDI and Export
13. La mobilité de la main-d'œuvre en Europe : le rôle des caractéristiques individuelles et de l'hétérogénéité entre pays
14. The Folklore of Sorting Algorithms
15. New issues in Indian macro policy.
16. Foreign direct investment in the Indian telecommunications sector
17. The name is absent
18. Prizes and Patents: Using Market Signals to Provide Incentives for Innovations
19. Synthesis and biological activity of α-galactosyl ceramide KRN7000 and galactosyl (α1→2) galactosyl ceramide
20. Sectoral specialisation in the EU a macroeconomic perspective