dated that renders the distinction between selectivity and specificity ultimately arbitrary. Such cases require the combination of selectivity and
specificity approaches. To accomplish this, the relevant natural vocalizations
are systematically decomposed into simpler signals. At the same time,
artificial sounds are used to synthesize increasingly closer approximations
(models) of the natural vocalizations. Optimally, both decomposition and
synthesis procedures are sensitive to behaviorally salient variation identified in the natural vocalizations. The goal is to bring the decomposed
and synthetic stimuli to some common intersection and thus to establish a
logical relationship between the variation in neuronal response and the
variation in the stimuli.
Combining the approaches of modified natural and artificial stimuli need
not result in convergence on the same solution—different parameter spaces
may be identified using the two approaches. In addition, errors are possible,
because there are practical limits to the size of a repertoire that depend on
the stability of the recordings and the natural rate of stimulus repetition. In
the absence of a specific model, these limits may prevent the choice of a
sufficient repertoire in cases of large and complex vocal repertoires or
highly selective neuronal responses. Errors in analysis may also result from
experimental decisions about the appropriate stimuli that have to be made
online without the benefit of retrospective analysis. These can be exceptionally challenging and exciting experiments! In the limited number of
cases where this approach has been employed, data from natural and
artificial stimuli have converged upon common solutions (e.g., Margoliash
1983). When the two approaches converge on a common set of parameters,
this gives confidence that a uniform model of the neuron’s response profile
has been achieved. In this case, the derived model can account for the
neuron’s response in terms of a specific set of acoustic parameters in the
natural vocalizations.
The success of a receptive-field model of neuronal response selectivity
can be independently determined by quantitative predictions of response
selectivity based on acoustic specificity that in turn can suggest specific
predictions regarding cellular mechanisms that give rise to the selective
responses. Achieving such quantitative predictions may itself require a significant modeling effort and is not frequently attempted. For example, the
responses to songs of some avian auditory thalamic neurons can be quantitatively predicted from their responses to tone and noise bursts (Bankes
and Margoliash 1993; Anderson et al. 1996). Achieving this result required
extensive testing to find the appropriate model (network) architectures.
Interestingly, the predictive power of the models was most sensitive to the
dynamics of the neuronal responses to the artificial stimuli. Model output
was less sensitive to manipulation of traditional static descriptions of neuronal response such as the frequency-amplitude response curves and rate
intensity functions.
342
T.Q. Gentner and D. Margoliash
specificity approaches. To accomplish this, the relevant natural vocalizations
are systematically decomposed into simpler signals. At the same time,
artificial sounds are used to synthesize increasingly closer approximations
(models) of the natural vocalizations. Optimally, both decomposition and
synthesis procedures are sensitive to behaviorally salient variation identified in the natural vocalizations. The goal is to bring the decomposed
and synthetic stimuli to some common intersection and thus to establish a
logical relationship between the variation in neuronal response and the
variation in the stimuli.
Combining the approaches of modified natural and artificial stimuli need
not result in convergence on the same solution—different parameter spaces
may be identified using the two approaches. In addition, errors are possible,
because there are practical limits to the size of a repertoire that depend on
the stability of the recordings and the natural rate of stimulus repetition. In
the absence of a specific model, these limits may prevent the choice of a
sufficient repertoire in cases of large and complex vocal repertoires or
highly selective neuronal responses. Errors in analysis may also result from
experimental decisions about the appropriate stimuli that have to be made
online without the benefit of retrospective analysis. These can be exceptionally challenging and exciting experiments! In the limited number of
cases where this approach has been employed, data from natural and
artificial stimuli have converged upon common solutions (e.g., Margoliash
1983). When the two approaches converge on a common set of parameters,
this gives confidence that a uniform model of the neuron’s response profile
has been achieved. In this case, the derived model can account for the
neuron’s response in terms of a specific set of acoustic parameters in the
natural vocalizations.
The success of a receptive-field model of neuronal response selectivity
can be independently determined by quantitative predictions of response
selectivity based on acoustic specificity that in turn can suggest specific
predictions regarding cellular mechanisms that give rise to the selective
responses. Achieving such quantitative predictions may itself require a significant modeling effort and is not frequently attempted. For example, the
responses to songs of some avian auditory thalamic neurons can be quantitatively predicted from their responses to tone and noise bursts (Bankes
and Margoliash 1993; Anderson et al. 1996). Achieving this result required
extensive testing to find the appropriate model (network) architectures.
Interestingly, the predictive power of the models was most sensitive to the
dynamics of the neuronal responses to the artificial stimuli. Model output
was less sensitive to manipulation of traditional static descriptions of neuronal response such as the frequency-amplitude response curves and rate
intensity functions.
342
T.Q. Gentner and D. Margoliash