Skip to main content
Have a personal or library account? Click to login
Computational Modeling of Contrast Sensitivity and Orientation Tuning in First-Episode and Chronic Schizophrenia Cover

Computational Modeling of Contrast Sensitivity and Orientation Tuning in First-Episode and Chronic Schizophrenia

Computational Psychiatry
Volume 1 (2017): Issue 0
By: Steven M. Silverstein,  Docia L. Demmin and  James A. Bednar  
Open Access
|Dec 2017
Figures & tables

Figures & Tables

Figure 1. 

Depiction of the sheets and connections in the gain control, adaptation, laterally connected (GCAL) class of models used in this study. Sheets include retina, lateral geniculate nucleus (LGN) On and LGN Off, and V1. Projections include afferent input to each of the LGN sheets and afferent input to V1 from each of the LGN sheets as well as lateral excitatory feedback within V1 (inner yellow circle) and a wider range of lateral inhibitory feedback within V1 (outer yellow circle). Examples of the training stimuli used in each model (i.e., pairs of oriented Gaussians) can be seen in the retinal sheet, with corresponding transformations in the LGN and V1 sheets.

Table 1. 

Summary of gain control, adaptation, laterally connected (GCAL) parameters manipulated in the schizophrenia models

Parameter manipulated Effect
Input strengthDecreased retinal input to LGN
Decreased LGN contrast gain control
Decreased LGN input to V1
Excitation and inhibitionIncreased lateral excitation within V1
Reduced lateral inhibition within V1
Reduced homeostatic adaptation rate within V1
PlasticityIncreased afferent learning rate at LGN-to-V1 connections
Increased excitatory learning rate within V1

[i] Note. LGN = lateral geniculate nucleus.

Figure 2. 

Stimuli used for posttraining testing. Left: low-spatial-frequency (SF) stimulus (frequency = 1.5 cycles per image). Right: medium-SF stimulus (frequency = 6 cycles per image).

Table 2. 

General classes of models run in the schizophrenia simulations and summary of the fit of models to chronic schizophrenia and/or first-episode schizophrenia data

Model Classes Result a
Increased V1 lateral excitation (range = 10% to 50%)None
Reduced V1 lateral inhibition (range = 10% to 50%)None
Increased V1 lateral excitation (20%) +
 Increased afferent learning rate at V1 (up to 3×)None
 or - Reduced V1 homeostatic adaptation rate (up to 90% ↓)None
 or - Increased V1 excitatory learning rate (0.0 to 0.2)None
Reduced V1 lateral inhibition (10%) +
 Increased afferent learning rate at V1 (up to 3×)Chronicb (best fit)
 or - Reduced V1 homeostatic adaptation rate (up to 90% ↓)Chronic (CS but not orientation tuning)
 or - Increased V1 excitatory learning rate (0.0 to 0.2)Chronic (second best fit)
Reduced retinal input to LGN and reduced LGN input to V1 (range = 10% to 50%) +FESc
 each of the manipulations described aboveNo significant improvementd over reduced retinal and LGN afferents alone
Reduced LGN contrast gain control (range = 25% to 75%)None

[i] Note. FES = first-episode schizophrenia. LGN = lateral geniculate nucleus.

[ii] a(i.e., fit for Chronic, FES, or none). bAt 10% reduced V1 lateral inhibition and 3× increased afferent learning rate at LGN synapses onto V1. cWith 15% reduction in both variables. dIncreasing the afferent learning rate on LGN synapses to V1 led to broadened orientation tuning (an issue that has not yet been studied in FES) while leaving increased CS for the LSF stimulus unaffected.

Figure 3. 

V1 orientation preference maps. A) The unmodified GCAL model after 20,000 iterations. B) The best fitting model for chronic schizophrenia data, involving 10% reduced V1 lateral inhibition and a tripled V1 afferent learning rate at iteration 10,001, with the model then run for another 10,000 iterations. C) The best fitting model for first-episode schizophrenia (FES) data, involving 15% reduced strength of retinal and LGN efferent activity at iteration 10,001, with the model then run for another 10,000 iterations. The qualitative similarity between the maps AC indicates that, despite perturbations to schizophrenia-relevant parameters that led to increases in contrast sensitivity (CS) (in C) and reduced CS and broadened orientation tuning (in B), realistic topography of orientation selectivity in V1 was maintained. D) Orientation key. Each color in the maps corresponds to maximum selectivity for the orientation denoted by the corresponding color in the key.

Figure 4. 

V1 retinotopic activation maps in response to the low-SF stimulus at 80% contrast. A) The stimulus, here compressed to show over a larger spatial extent, and so SF appears higher than in the actual stimulus, shown in Figure 2. B) Unmodified GCAL model after 20,000 iterations (mean activation = 0.029; max. = 0.512). C) The best fitting model for chronic schizophrenia data (mean = 0.007; max. = 0.386). D) The best fitting model for FES data (mean = 0.041, max. = 0.540). Greater activation is indicated by increased brightness. Reduced activation to the low-SF stimulus can be seen in C, the chronic schizophrenia model. Increased activation to the same stimulus can be seen in D, in the FES model. These changes, relative to the unmodified model (B), and differences from each other are consistent with published data on CS in these groups.

Figure 5. 

V1 retinotopic activation maps showing response to the medium-SF stimulus at 80% contrast. A) The unmodified GCAL model after 20,000 iterations (mean activation 0.137; max. = 0.587). B) The best fitting model for chronic schizophrenia data (mean activation = 0.140, max. = 0.550). C) The best fitting model for FES data (mean activation = 0.131; max. = 0.564). Note that, as predicted, at this higher SF, activation is higher than shown in Figure 4. Also note that a greater spread of activation can be seen in the chronic schizophrenia model compared to the other models (i.e., there is activation of orientation-selective simple cells that are not activated in the other models). This greater activation of orientation-selective cells that are normally relatively silent may be the basis of broadened orientation tuning in this group.

Figure 6. 

Combined orientation preference and activation maps showing response to the medium-SF stimulus at 80% contrast. A) The unmodified GCAL model after 20,000 iterations. B) The best fitting model for chronic schizophrenia data. C) The best fitting model for FES data. It can be seen here that the broader orientation tuning in the chronic schizophrenia model (B; see Figure 5) arises owing to greater activation of cells not selective for a vertical orientation (see D for orientation key).

Figure 7. 

Orientation tuning histograms (V1 orientation map weighted by the strength of activation of each unit) showing responses (y axis, in arbitrary units) to the medium-SF stimulus at 80% contrast. A) The unmodified GCAL model after 20,000 iterations. B) The best fitting model for chronic schizophrenia data. C) The best fitting model for FES data. The x axis depicts orientation in radians. The unmodified model (A) shows a typical Gaussian curve with the peak orientation at vertical (90° or π/2 or ∼1.57 rad). In the chronic schizophrenia model (B), there is suppression of cells signaling this orientation and a near to bimodal distribution with separated peaks at both lesser and greater orientations, which may be the basis for broadened orientation tuning in this population. The high excess kurtosis value for this panel indicates that the chronic model is also associated with the greatest activation in neurons signaling orientations far from the peak. In the FES model (C), the peak response is to a vertical stimulus, and the shape of the distribution is essentially Gaussian. Note that in models B and C, there is slightly greater peak activation for the peak response compared to model A, as indicated by the larger range on the y axis.

Figure 8. 

Activation (y axis, in arbitrary units) values across contrast levels (x axis) for the unmodified GCAL model, the best fitting chronic schizophrenia model, and the best fitting FES model. A) The low-SF stimulus. B) The medium-SF stimulus. In both panels, the unmodified model is represented by the thick blue line. The pattern displayed here is one of (a) increased activation in the FES model, and reduced activation in the chronic schizophrenia model, in response to the low-SF stimulus and (b) normal activation in the FES model, and slightly reduced activation at medium contrast levels in the chronic schizophrenia model, in response to the medium-SF stimulus, consistent with published observations for these patient groups. Note that the activation range is higher for the medium-SF stimulus, as indicated by the y axis in B.

Figure 9. 

Activation values (y axis, in arbitrary units) across contrast levels for the unmodified GCAL model, after 20,000 iterations, and several alternative schizophrenia models discussed in the text. Left: activity in response to the low-SF stimulus. Right: activity in response to the medium-SF stimulus. Model A (unmodified), represented by the thick blue line, is the basic GCAL model. All other models represent cases with perturbations at the 10,001st iteration, as follows: Model B, 25% increase in V1 lateral excitation; Model C, 25% reduction in V1 lateral inhibition; Model D, 20% increase in V1 lateral excitation and an increase in the V1 excitatory learning rate to 0.01; Model E, 10% reduction in V1 lateral inhibition and an increase in the V1 excitatory learning rate to 0.01; Model F (barely visible owing to almost complete overlap with Model A, but indicated by an asterisk), an increase in the V1 afferent learning rate to 0.03 without any change to lateral inhibition; Model G, 15% reduction in both retinal and LGN efferents in combination with an increase in the V1 afferent learning rate to 0.03. Note that although Models G in the left panel and F in the right panel are reasonable fits to previously published FES CS data, and although Models E in the left panel and C and D in the right panel are reasonable fits to previously published chronic schizophrenia data, none of these models are a good fit across both the left and right panels. Results for these same models are also depicted in Figure 10.

Figure 10. 

Orientation tuning histograms showing response to the medium-SF stimulus at 80% contrast for Models A-G (as described in Figure 9). Although Models DG would be expected to be associated with broadened orientation tuning, none of these models were also associated with activation data that match what would be expected based on published data on CS from chronic or FE schizophrenia.

DOI: https://doi.org/10.1162/CPSY_a_00005 | Journal eISSN: 2379-6227
Journal RSS Feed
Language: English
Submitted on: Jan 17, 2017
Accepted on: May 16, 2017
Published on: Dec 1, 2017
Published by: MIT Press
In partnership with: Paradigm Publishing Services
Publication frequency: 1 issue per year
Keywords:
schizophrenia,
vision,
perception,
contrast sensitivity,
orientation tuning,
modeling,
first episode,
Hebbian learning,
gain control,
inhibition,
excitation

© 2017 Steven M. Silverstein, Docia L. Demmin, James A. Bednar, published by MIT Press
This work is licensed under the Creative Commons Attribution 4.0 License.

Volume 1 (2017): Issue 0
Previous article