Limits...
Differing effects of attention in single-units and populations are well predicted by heterogeneous tuning and the normalization model of attention.

Hara Y, Pestilli F, Gardner JL - Front Comput Neurosci (2014)

Bottom Line: The normalization model of attention elegantly predicts the diversity of effects of attention reported in single-units well-tuned to the stimulus, but what predictions does it make for more realistic populations of neurons with heterogeneous tuning?We found that within the population, neurons well-tuned to the stimulus showed a response-gain effect, while less-well-tuned neurons showed a contrast-gain effect.More generally, computational models can unify physiological findings across different scales of measurement, and make links to behavior, but only if factors such as heterogeneous tuning within a population are properly accounted for.

View Article: PubMed Central - PubMed

Affiliation: Laboratory for Human Systems Neuroscience, RIKEN Brain Science Institute Wako, Japan.

ABSTRACT
Single-unit measurements have reported many different effects of attention on contrast-response (e.g., contrast-gain, response-gain, additive-offset dependent on visibility), while functional imaging measurements have more uniformly reported increases in response across all contrasts (additive-offset). The normalization model of attention elegantly predicts the diversity of effects of attention reported in single-units well-tuned to the stimulus, but what predictions does it make for more realistic populations of neurons with heterogeneous tuning? Are predictions in accordance with population-scale measurements? We used functional imaging data from humans to determine a realistic ratio of attention-field to stimulus-drive size (a key parameter for the model) and predicted effects of attention in a population of model neurons with heterogeneous tuning. We found that within the population, neurons well-tuned to the stimulus showed a response-gain effect, while less-well-tuned neurons showed a contrast-gain effect. Averaged across the population, these disparate effects of attention gave rise to additive-offsets in contrast-response, similar to reports in human functional imaging as well as population averages of single-units. Differences in predictions for single-units and populations were observed across a wide range of model parameters (ratios of attention-field to stimulus-drive size and the amount of baseline response modifiable by attention), offering an explanation for disparity in physiological reports. Thus, by accounting for heterogeneity in tuning of realistic neuronal populations, the normalization model of attention can not only predict responses of well-tuned neurons, but also the activity of large populations of neurons. More generally, computational models can unify physiological findings across different scales of measurement, and make links to behavior, but only if factors such as heterogeneous tuning within a population are properly accounted for.

No MeSH data available.


Group analyses of spatial characteristics of task-related signals. Gaussians were fit (bootstrap CI 95%) to the mean across subjects of contrast-sensitivity (A), cue-sensitivity (B), and BOLD amplitudes (C) binned by eccentricity. An eccentricity of 0° represents the center of the response to the stimulus as determined by the localizer. The shade of each point reflects the reliability of the eccentricity measured for that bin (coherence threshold of 0.4). In (C), a sum of two Gaussians was used to account for the large activity near fixation (yellow dashed line) and the activity in response to the contrast grating (solid blue). (D) All fits were normalized and superimposed to visualize the range of each effect. The dark gray shaded area corresponds to the spatial extent of the contrast grating, and the light gray shaded area represents the subject's viewing range in the scanner.
© Copyright Policy - open-access
Related In: Results  -  Collection

License
getmorefigures.php?uid=PMC3928538&req=5

Figure 3: Group analyses of spatial characteristics of task-related signals. Gaussians were fit (bootstrap CI 95%) to the mean across subjects of contrast-sensitivity (A), cue-sensitivity (B), and BOLD amplitudes (C) binned by eccentricity. An eccentricity of 0° represents the center of the response to the stimulus as determined by the localizer. The shade of each point reflects the reliability of the eccentricity measured for that bin (coherence threshold of 0.4). In (C), a sum of two Gaussians was used to account for the large activity near fixation (yellow dashed line) and the activity in response to the contrast grating (solid blue). (D) All fits were normalized and superimposed to visualize the range of each effect. The dark gray shaded area corresponds to the spatial extent of the contrast grating, and the light gray shaded area represents the subject's viewing range in the scanner.

Mentions: The three attributes were averaged across subjects to analyze the mean extent of their effects as a function of radial eccentricity from the grating center. The grating eccentricity was estimated by averaging the eccentricities of voxels in V1, V2, and V3 whose localizer coherence exceeded 0.5. Each voxel was binned into one of seventeen eccentricity bins centered at the grating (Figure 3, 0° visual angle). The average BOLD response amplitude, contrast-sensitivity, and cue-sensitivity were calculated for each bin, resulting in sensitivity profiles as a function of radial distance from the grating center. These profiles were averaged across subjects, then fitted: the contrast- and cue-sensitivity profiles were fit with a single Gaussian (Figures 3A,B; mean fixed at 0° visual angle, offset fixed at 0, σ and amplitude were fit parameters), and the BOLD response amplitude profile was fit by a sum of two Gaussians, the first Gaussian accounting for the large response seen near the fovea and the second Gaussian accounting for the response evoked by the stimulus grating (Figure 3C; Mean of first Gaussian constrained to be < −2° visual angle, i.e., foveal to the grating location. Mean of second Gaussian fixed at 0° visual angle. Offsets, σ and amplitudes for both Gaussians were fit parameters).


Differing effects of attention in single-units and populations are well predicted by heterogeneous tuning and the normalization model of attention.

Hara Y, Pestilli F, Gardner JL - Front Comput Neurosci (2014)

Group analyses of spatial characteristics of task-related signals. Gaussians were fit (bootstrap CI 95%) to the mean across subjects of contrast-sensitivity (A), cue-sensitivity (B), and BOLD amplitudes (C) binned by eccentricity. An eccentricity of 0° represents the center of the response to the stimulus as determined by the localizer. The shade of each point reflects the reliability of the eccentricity measured for that bin (coherence threshold of 0.4). In (C), a sum of two Gaussians was used to account for the large activity near fixation (yellow dashed line) and the activity in response to the contrast grating (solid blue). (D) All fits were normalized and superimposed to visualize the range of each effect. The dark gray shaded area corresponds to the spatial extent of the contrast grating, and the light gray shaded area represents the subject's viewing range in the scanner.
© Copyright Policy - open-access
Related In: Results  -  Collection

License
Show All Figures
getmorefigures.php?uid=PMC3928538&req=5

Figure 3: Group analyses of spatial characteristics of task-related signals. Gaussians were fit (bootstrap CI 95%) to the mean across subjects of contrast-sensitivity (A), cue-sensitivity (B), and BOLD amplitudes (C) binned by eccentricity. An eccentricity of 0° represents the center of the response to the stimulus as determined by the localizer. The shade of each point reflects the reliability of the eccentricity measured for that bin (coherence threshold of 0.4). In (C), a sum of two Gaussians was used to account for the large activity near fixation (yellow dashed line) and the activity in response to the contrast grating (solid blue). (D) All fits were normalized and superimposed to visualize the range of each effect. The dark gray shaded area corresponds to the spatial extent of the contrast grating, and the light gray shaded area represents the subject's viewing range in the scanner.
Mentions: The three attributes were averaged across subjects to analyze the mean extent of their effects as a function of radial eccentricity from the grating center. The grating eccentricity was estimated by averaging the eccentricities of voxels in V1, V2, and V3 whose localizer coherence exceeded 0.5. Each voxel was binned into one of seventeen eccentricity bins centered at the grating (Figure 3, 0° visual angle). The average BOLD response amplitude, contrast-sensitivity, and cue-sensitivity were calculated for each bin, resulting in sensitivity profiles as a function of radial distance from the grating center. These profiles were averaged across subjects, then fitted: the contrast- and cue-sensitivity profiles were fit with a single Gaussian (Figures 3A,B; mean fixed at 0° visual angle, offset fixed at 0, σ and amplitude were fit parameters), and the BOLD response amplitude profile was fit by a sum of two Gaussians, the first Gaussian accounting for the large response seen near the fovea and the second Gaussian accounting for the response evoked by the stimulus grating (Figure 3C; Mean of first Gaussian constrained to be < −2° visual angle, i.e., foveal to the grating location. Mean of second Gaussian fixed at 0° visual angle. Offsets, σ and amplitudes for both Gaussians were fit parameters).

Bottom Line: The normalization model of attention elegantly predicts the diversity of effects of attention reported in single-units well-tuned to the stimulus, but what predictions does it make for more realistic populations of neurons with heterogeneous tuning?We found that within the population, neurons well-tuned to the stimulus showed a response-gain effect, while less-well-tuned neurons showed a contrast-gain effect.More generally, computational models can unify physiological findings across different scales of measurement, and make links to behavior, but only if factors such as heterogeneous tuning within a population are properly accounted for.

View Article: PubMed Central - PubMed

Affiliation: Laboratory for Human Systems Neuroscience, RIKEN Brain Science Institute Wako, Japan.

ABSTRACT
Single-unit measurements have reported many different effects of attention on contrast-response (e.g., contrast-gain, response-gain, additive-offset dependent on visibility), while functional imaging measurements have more uniformly reported increases in response across all contrasts (additive-offset). The normalization model of attention elegantly predicts the diversity of effects of attention reported in single-units well-tuned to the stimulus, but what predictions does it make for more realistic populations of neurons with heterogeneous tuning? Are predictions in accordance with population-scale measurements? We used functional imaging data from humans to determine a realistic ratio of attention-field to stimulus-drive size (a key parameter for the model) and predicted effects of attention in a population of model neurons with heterogeneous tuning. We found that within the population, neurons well-tuned to the stimulus showed a response-gain effect, while less-well-tuned neurons showed a contrast-gain effect. Averaged across the population, these disparate effects of attention gave rise to additive-offsets in contrast-response, similar to reports in human functional imaging as well as population averages of single-units. Differences in predictions for single-units and populations were observed across a wide range of model parameters (ratios of attention-field to stimulus-drive size and the amount of baseline response modifiable by attention), offering an explanation for disparity in physiological reports. Thus, by accounting for heterogeneity in tuning of realistic neuronal populations, the normalization model of attention can not only predict responses of well-tuned neurons, but also the activity of large populations of neurons. More generally, computational models can unify physiological findings across different scales of measurement, and make links to behavior, but only if factors such as heterogeneous tuning within a population are properly accounted for.

No MeSH data available.