Limits...
Cortical plasticity as a mechanism for storing Bayesian priors in sensory perception.

Köver H, Bao S - PLoS ONE (2010)

Bottom Line: Human perception of ambiguous sensory signals is biased by prior experiences.It is not known how such prior information is encoded, retrieved and combined with sensory information by neurons.For the case of auditory perception, we use a computational model to show that prior information about sound frequency distributions may be stored in the size of primary auditory cortex frequency representations, read-out by elevated baseline activity in all neurons and combined with sensory-evoked activity to generate a perception that conforms to Bayesian integration theory.

View Article: PubMed Central - PubMed

Affiliation: Helen Wills Neuroscience Institute, University of California, Berkeley, California, United States of America.

ABSTRACT
Human perception of ambiguous sensory signals is biased by prior experiences. It is not known how such prior information is encoded, retrieved and combined with sensory information by neurons. Previous authors have suggested dynamic encoding mechanisms for prior information, whereby top-down modulation of firing patterns on a trial-by-trial basis creates short-term representations of priors. Although such a mechanism may well account for perceptual bias arising in the short-term, it does not account for the often irreversible and robust changes in perception that result from long-term, developmental experience. Based on the finding that more frequently experienced stimuli gain greater representations in sensory cortices during development, we reasoned that prior information could be stored in the size of cortical sensory representations. For the case of auditory perception, we use a computational model to show that prior information about sound frequency distributions may be stored in the size of primary auditory cortex frequency representations, read-out by elevated baseline activity in all neurons and combined with sensory-evoked activity to generate a perception that conforms to Bayesian integration theory. Our results suggest an alternative neural mechanism for experience-induced long-term perceptual bias in the context of auditory perception. They make the testable prediction that the extent of such perceptual prior bias is modulated by both the degree of cortical reorganization and the magnitude of spontaneous activity in primary auditory cortex. Given that cortical over-representation of frequently experienced stimuli, as well as perceptual bias towards such stimuli is a common phenomenon across sensory modalities, our model may generalize to sensory perception, rather than being specific to auditory perception.

Show MeSH

Related in: MedlinePlus

Neuronal population activity and derived log-likelihood functions.Left panels show population activity of the model 7-kHz-over-represented AI, and the right panels show stimulus log-likelihood functions. (a) Response of the model to a 4-kHz tone pip (b) Elevated baseline activity in the absence of a stimulus (c) Summed spontaneous and 4-kHz-evoked activity). Each bar in the left panels represents the firing rate of a model neuron. The neurons are arranged by characteristic frequency, with low frequency-tuned neurons on the left and high frequency-tuned neurons on the right. Blue dotted lines in the right panels show the input frequency, red dotted lines show the over-represented frequency, and the black dotted lines mark the peaks of the log-likelihood functions.
© Copyright Policy
Related In: Results  -  Collection


getmorefigures.php?uid=PMC2864750&req=5

pone-0010497-g002: Neuronal population activity and derived log-likelihood functions.Left panels show population activity of the model 7-kHz-over-represented AI, and the right panels show stimulus log-likelihood functions. (a) Response of the model to a 4-kHz tone pip (b) Elevated baseline activity in the absence of a stimulus (c) Summed spontaneous and 4-kHz-evoked activity). Each bar in the left panels represents the firing rate of a model neuron. The neurons are arranged by characteristic frequency, with low frequency-tuned neurons on the left and high frequency-tuned neurons on the right. Blue dotted lines in the right panels show the input frequency, red dotted lines show the over-represented frequency, and the black dotted lines mark the peaks of the log-likelihood functions.

Mentions: We therefore modeled Bayesian integration of sensory input and prior-based expectation by calculating the stimulus likelihood function derived from the linear superposition of stimulus-evoked activity and elevated spontaneous activity ( and )(Fig. 2c).(5)When the frequency representation is homogeneous, equation 5 may be simplified as,(6)which is in the form of Bayes rule. With inhomogeneous frequency representations, there is a small deviation from Bayes rule caused by an additional term, (see equation 5).


Cortical plasticity as a mechanism for storing Bayesian priors in sensory perception.

Köver H, Bao S - PLoS ONE (2010)

Neuronal population activity and derived log-likelihood functions.Left panels show population activity of the model 7-kHz-over-represented AI, and the right panels show stimulus log-likelihood functions. (a) Response of the model to a 4-kHz tone pip (b) Elevated baseline activity in the absence of a stimulus (c) Summed spontaneous and 4-kHz-evoked activity). Each bar in the left panels represents the firing rate of a model neuron. The neurons are arranged by characteristic frequency, with low frequency-tuned neurons on the left and high frequency-tuned neurons on the right. Blue dotted lines in the right panels show the input frequency, red dotted lines show the over-represented frequency, and the black dotted lines mark the peaks of the log-likelihood functions.
© Copyright Policy
Related In: Results  -  Collection

Show All Figures
getmorefigures.php?uid=PMC2864750&req=5

pone-0010497-g002: Neuronal population activity and derived log-likelihood functions.Left panels show population activity of the model 7-kHz-over-represented AI, and the right panels show stimulus log-likelihood functions. (a) Response of the model to a 4-kHz tone pip (b) Elevated baseline activity in the absence of a stimulus (c) Summed spontaneous and 4-kHz-evoked activity). Each bar in the left panels represents the firing rate of a model neuron. The neurons are arranged by characteristic frequency, with low frequency-tuned neurons on the left and high frequency-tuned neurons on the right. Blue dotted lines in the right panels show the input frequency, red dotted lines show the over-represented frequency, and the black dotted lines mark the peaks of the log-likelihood functions.
Mentions: We therefore modeled Bayesian integration of sensory input and prior-based expectation by calculating the stimulus likelihood function derived from the linear superposition of stimulus-evoked activity and elevated spontaneous activity ( and )(Fig. 2c).(5)When the frequency representation is homogeneous, equation 5 may be simplified as,(6)which is in the form of Bayes rule. With inhomogeneous frequency representations, there is a small deviation from Bayes rule caused by an additional term, (see equation 5).

Bottom Line: Human perception of ambiguous sensory signals is biased by prior experiences.It is not known how such prior information is encoded, retrieved and combined with sensory information by neurons.For the case of auditory perception, we use a computational model to show that prior information about sound frequency distributions may be stored in the size of primary auditory cortex frequency representations, read-out by elevated baseline activity in all neurons and combined with sensory-evoked activity to generate a perception that conforms to Bayesian integration theory.

View Article: PubMed Central - PubMed

Affiliation: Helen Wills Neuroscience Institute, University of California, Berkeley, California, United States of America.

ABSTRACT
Human perception of ambiguous sensory signals is biased by prior experiences. It is not known how such prior information is encoded, retrieved and combined with sensory information by neurons. Previous authors have suggested dynamic encoding mechanisms for prior information, whereby top-down modulation of firing patterns on a trial-by-trial basis creates short-term representations of priors. Although such a mechanism may well account for perceptual bias arising in the short-term, it does not account for the often irreversible and robust changes in perception that result from long-term, developmental experience. Based on the finding that more frequently experienced stimuli gain greater representations in sensory cortices during development, we reasoned that prior information could be stored in the size of cortical sensory representations. For the case of auditory perception, we use a computational model to show that prior information about sound frequency distributions may be stored in the size of primary auditory cortex frequency representations, read-out by elevated baseline activity in all neurons and combined with sensory-evoked activity to generate a perception that conforms to Bayesian integration theory. Our results suggest an alternative neural mechanism for experience-induced long-term perceptual bias in the context of auditory perception. They make the testable prediction that the extent of such perceptual prior bias is modulated by both the degree of cortical reorganization and the magnitude of spontaneous activity in primary auditory cortex. Given that cortical over-representation of frequently experienced stimuli, as well as perceptual bias towards such stimuli is a common phenomenon across sensory modalities, our model may generalize to sensory perception, rather than being specific to auditory perception.

Show MeSH
Related in: MedlinePlus