employing to achieve signal recognition, they were producing similar
response biases.
In a subsequent study, the authors examined the effect of history on
response biases. Ryan and Rand (1995, 1999b) had used phylogenetic techniques to estimate what the calls of ancestors of túngara frogs might have
sounded like. The correlation between evolutionary relationship (estimated
as similarity in DNA sequences) and call similarity was not statistically significant—calls of close relatives were not more likely to sound alike than
calls of more distant relatives. Nevertheless, female responses to the calls
were predicted by phylogenetic relationship as well as, and independent
from, overall call similarity. These results suggested that the history of the
receiver can influence its response biases.
Phelps and Ryan (2000) trained neural networks along three distinct histories. In the first, the mimetic history, networks were trained to the call at
the root of the phylogenetic tree (Fig. 5.10). Once the networks reached
recognition criteria, they were trained to the call that was the next most
immediate ancestor to the túngara frog. This procedure was continued on
the line of descent to the túngara frogs until the networks were trained to
the túngara frog call itself. Thus, these networks had a history of first being
trained to the calls of the three direct ancestors of túngara frogs before
being trained to the target call, the túngara frog call. This was replicated 20
times, and the most discriminating net in each replicate was later tested.
The authors conducted the same procedure for two control evolutionary
histories. For one, the random history, three calls were picked at random
from the sample of heterospecific and ancestral calls. The nets were trained
to these calls prior to being trained to the túngara frog calls. There were 20
random histories, and the most discriminating net in each was later tested.
In another control, the calls used in the mimetic history were rotated 180°
in principal component space and the new calls synthesized.The path length
among these “mirror” calls was identical to the path length among the calls
in the mimetic history; both of these path lengths were longer than that of
the random histories. There were 20 replicates and, as with the other histories, the most discriminating net in each was saved.
The best nets from the mimetic, random, and mirror histories were tested
against the same set of heterospecific and ancestral calls as with the ahistoric nets. Only the networks that were trained along the mimetic history
significantly predicted the response biases of the túngara frog females.
These results suggest that the past history of tasks a receiver needs to
accomplish influences the computational strategies it uses to accomplish
current tasks.
6.5. Summary and Conclusions
Long-distance acoustic signals are prevalent as mate-recognition signals
in a number of diverse taxa and are accessible for studies by behaviorists,
264
M.J. Ryan and N.M. Kime
response biases.
In a subsequent study, the authors examined the effect of history on
response biases. Ryan and Rand (1995, 1999b) had used phylogenetic techniques to estimate what the calls of ancestors of túngara frogs might have
sounded like. The correlation between evolutionary relationship (estimated
as similarity in DNA sequences) and call similarity was not statistically significant—calls of close relatives were not more likely to sound alike than
calls of more distant relatives. Nevertheless, female responses to the calls
were predicted by phylogenetic relationship as well as, and independent
from, overall call similarity. These results suggested that the history of the
receiver can influence its response biases.
Phelps and Ryan (2000) trained neural networks along three distinct histories. In the first, the mimetic history, networks were trained to the call at
the root of the phylogenetic tree (Fig. 5.10). Once the networks reached
recognition criteria, they were trained to the call that was the next most
immediate ancestor to the túngara frog. This procedure was continued on
the line of descent to the túngara frogs until the networks were trained to
the túngara frog call itself. Thus, these networks had a history of first being
trained to the calls of the three direct ancestors of túngara frogs before
being trained to the target call, the túngara frog call. This was replicated 20
times, and the most discriminating net in each replicate was later tested.
The authors conducted the same procedure for two control evolutionary
histories. For one, the random history, three calls were picked at random
from the sample of heterospecific and ancestral calls. The nets were trained
to these calls prior to being trained to the túngara frog calls. There were 20
random histories, and the most discriminating net in each was later tested.
In another control, the calls used in the mimetic history were rotated 180°
in principal component space and the new calls synthesized.The path length
among these “mirror” calls was identical to the path length among the calls
in the mimetic history; both of these path lengths were longer than that of
the random histories. There were 20 replicates and, as with the other histories, the most discriminating net in each was saved.
The best nets from the mimetic, random, and mirror histories were tested
against the same set of heterospecific and ancestral calls as with the ahistoric nets. Only the networks that were trained along the mimetic history
significantly predicted the response biases of the túngara frog females.
These results suggest that the past history of tasks a receiver needs to
accomplish influences the computational strategies it uses to accomplish
current tasks.
6.5. Summary and Conclusions
Long-distance acoustic signals are prevalent as mate-recognition signals
in a number of diverse taxa and are accessible for studies by behaviorists,
264
M.J. Ryan and N.M. Kime