FaceSim – Results¶
Notebook to summarize results of the FaceSim study.
Author: Simon M. Hofmann
Years: 2023-2024Results¶
Number of participants that took part in session 2D: 1530 (n_female = 714, n_male = 716, n_unknown = 100) Age of participants in session 2D: range = 19 - 65, mean = 31.27 ± 10.75 SD years Number of participants that completed the task in session 2D: 1397 Number of participants that failed the attention checks in session 2D: 104 Number of participants that entered NO CODE in session 2D: 21 Number of participants that entered unknown code in session 2D: 8 *-*-*-*-*-*-*-*-*-*-*-*-*-*-*-*-*-*-*-*-*-*-*-*-*-*-*-*-*-*-*-*-*-*-*-*-*-*-*-*-* Number of participants that took part in session 3D: 1414 (n_female = 667, n_male = 662, n_unknown = 85) Age of participants in session 3D: range = 18 - 65, mean = 32.57 ± 11.24 SD years Number of participants that completed the task in session 3D: 1313 Number of participants that failed the attention checks in session 3D: 78 Number of participants that entered NO CODE in session 3D: 22 Number of participants that entered unknown code in session 3D: 1 *-*-*-*-*-*-*-*-*-*-*-*-*-*-*-*-*-*-*-*-*-*-*-*-*-*-*-*-*-*-*-*-*-*-*-*-*-*-*-*-* Age of all participants: range = 18 - 65, mean = 31.90 ± 11.01 SD years
Trials excluded¶
Pprobabity being caught < 3 times across the whole experiment The chance of having less than 3 catch(es) in 9 catch-trials is: 0.83% Probabity being not caught in one block The chance of having less than 1 catch(es) in 3 catch-trials is: 3.70%
2D Session: In session 2D 18.5% of trials were removed. 3D Session: In session 3D 16.9% of trials were removed.
Response times¶
Response time for session 2D:
count 194261.000000
mean 3.479793
std 1.498887
min 0.651300
25% 2.374800
50% 3.098500
75% 4.214500
max 9.999999
Name: response_time, dtype: float64
Response time for session 3D:
count 185473.000000
mean 3.543740
std 1.496115
min 0.633200
25% 2.433000
50% 3.169599
75% 4.299999
max 9.999999
Name: response_time, dtype: float64
ttest(rt_2d, rt_3d): t-statistic: -13.15, p-value <= 1.66e-39
Cohen's d: -0.043
Generalized Linear Model Regression Results
==============================================================================
Dep. Variable: response_time No. Observations: 379734
Model: GLM Df Residuals: 379732
Model Family: InverseGaussian Df Model: 1
Link Function: identity Scale: 0.051843
Method: IRLS Log-Likelihood: -6.3403e+05
Date: Thu, 21 Nov 2024 Deviance: 18458.
Time: 15:01:50 Pearson chi2: 1.97e+04
No. Iterations: 5 Pseudo R-squ. (CS): 0.0004551
Covariance Type: nonrobust
====================================================================================
coef std err z P>|z| [0.025 0.975]
------------------------------------------------------------------------------------
Intercept 3.4798 0.003 1037.700 0.000 3.473 3.486
C(session)[T.3D] 0.0639 0.005 13.140 0.000 0.054 0.073
====================================================================================
Correlate response times of triplet ID's between sessions¶
Response times as function of similarity¶
pearsonr(sim ~ RT) (order=1): PearsonRResult(statistic=0.23160084101455494, pvalue=2.8385370557883092e-61)
pearsonr(sim ~ RT) (order=2): PearsonRResult(statistic=0.22948643784766498, pvalue=3.658911721046632e-60)
pearsonr(sim ~ RT) (order=1): PearsonRResult(statistic=0.26796668659180756, pvalue=3.6611021617578353e-82)
pearsonr(sim ~ RT) (order=2): PearsonRResult(statistic=0.2649071451455536, pvalue=2.845448723742933e-80)
2D
OLS Regression Results
==============================================================================
Dep. Variable: response_time R-squared: 0.054
Model: OLS Adj. R-squared: 0.053
Method: Least Squares F-statistic: 280.4
Date: Thu, 21 Nov 2024 Prob (F-statistic): 2.84e-61
Time: 15:02:07 Log-Likelihood: 1586.2
No. Observations: 4950 AIC: -3168.
Df Residuals: 4948 BIC: -3155.
Df Model: 1
Covariance Type: nonrobust
==============================================================================
coef std err t P>|t| [0.025 0.975]
------------------------------------------------------------------------------
const 3.3997 0.005 627.696 0.000 3.389 3.410
x1 0.2415 0.014 16.747 0.000 0.213 0.270
==============================================================================
Omnibus: 10.620 Durbin-Watson: 1.836
Prob(Omnibus): 0.005 Jarque-Bera (JB): 12.597
Skew: -0.031 Prob(JB): 0.00184
Kurtosis: 3.239 Cond. No. 6.43
==============================================================================
Notes:
[1] Standard Errors assume that the covariance matrix of the errors is correctly specified.
(2D-session: 1-order polynomial fit: R2=0.054; b0=3.400, SE0=0.005, t0=627.696, p0<0.000; b1=0.241, SE1=0.014, t1=16.747, p1<0.000; AIC=-3168, BIC=-3155).
OLS Regression Results
==============================================================================
Dep. Variable: response_time R-squared: 0.229
Model: OLS Adj. R-squared: 0.229
Method: Least Squares F-statistic: 735.5
Date: Thu, 21 Nov 2024 Prob (F-statistic): 2.21e-280
Time: 15:02:07 Log-Likelihood: 2094.0
No. Observations: 4950 AIC: -4182.
Df Residuals: 4947 BIC: -4163.
Df Model: 2
Covariance Type: nonrobust
==============================================================================
coef std err t P>|t| [0.025 0.975]
------------------------------------------------------------------------------
const 3.1478 0.009 351.479 0.000 3.130 3.165
x1 1.9309 0.052 37.144 0.000 1.829 2.033
x2 -2.2061 0.066 -33.568 0.000 -2.335 -2.077
==============================================================================
Omnibus: 27.794 Durbin-Watson: 1.913
Prob(Omnibus): 0.000 Jarque-Bera (JB): 28.107
Skew: 0.183 Prob(JB): 7.88e-07
Kurtosis: 3.049 Cond. No. 39.5
==============================================================================
Notes:
[1] Standard Errors assume that the covariance matrix of the errors is correctly specified.
(2D-session: 2-order polynomial fit: R2=0.229; b0=3.148, SE0=0.009, t0=351.479, p0<0.000; b1=1.931, SE1=0.052, t1=37.144, p1<0.000; b2=-2.206, SE2=0.066, t2=-33.568, p2<0.000; AIC=-4182, BIC=-4163).
3D
OLS Regression Results
==============================================================================
Dep. Variable: response_time R-squared: 0.072
Model: OLS Adj. R-squared: 0.072
Method: Least Squares F-statistic: 382.8
Date: Thu, 21 Nov 2024 Prob (F-statistic): 3.66e-82
Time: 15:02:07 Log-Likelihood: 1423.8
No. Observations: 4950 AIC: -2844.
Df Residuals: 4948 BIC: -2831.
Df Model: 1
Covariance Type: nonrobust
==============================================================================
coef std err t P>|t| [0.025 0.975]
------------------------------------------------------------------------------
const 3.4475 0.006 622.442 0.000 3.437 3.458
x1 0.2877 0.015 19.565 0.000 0.259 0.316
==============================================================================
Omnibus: 38.559 Durbin-Watson: 1.788
Prob(Omnibus): 0.000 Jarque-Bera (JB): 41.096
Skew: -0.186 Prob(JB): 1.19e-09
Kurtosis: 3.245 Cond. No. 6.35
==============================================================================
Notes:
[1] Standard Errors assume that the covariance matrix of the errors is correctly specified.
(3D-session: 1-order polynomial fit: R2=0.072; b0=3.448, SE0=0.006, t0=622.442, p0<0.000; b1=0.288, SE1=0.015, t1=19.565, p1<0.000; AIC=-2844, BIC=-2831).
OLS Regression Results
==============================================================================
Dep. Variable: response_time R-squared: 0.255
Model: OLS Adj. R-squared: 0.255
Method: Least Squares F-statistic: 846.9
Date: Thu, 21 Nov 2024 Prob (F-statistic): 4.75e-317
Time: 15:02:07 Log-Likelihood: 1968.1
No. Observations: 4950 AIC: -3930.
Df Residuals: 4947 BIC: -3911.
Df Model: 2
Covariance Type: nonrobust
==============================================================================
coef std err t P>|t| [0.025 0.975]
------------------------------------------------------------------------------
const 3.1860 0.009 354.354 0.000 3.168 3.204
x1 2.0575 0.052 39.255 0.000 1.955 2.160
x2 -2.3142 0.066 -34.886 0.000 -2.444 -2.184
==============================================================================
Omnibus: 0.521 Durbin-Watson: 1.854
Prob(Omnibus): 0.771 Jarque-Bera (JB): 0.473
Skew: 0.016 Prob(JB): 0.790
Kurtosis: 3.036 Cond. No. 38.9
==============================================================================
Notes:
[1] Standard Errors assume that the covariance matrix of the errors is correctly specified.
(3D-session: 2-order polynomial fit: R2=0.255; b0=3.186, SE0=0.009, t0=354.354, p0<0.000; b1=2.057, SE1=0.052, t1=39.255, p1<0.000; b2=-2.314, SE2=0.066, t2=-34.886, p2<0.000; AIC=-3930, BIC=-3911).
Predicting human similarity judgments¶
Across all samples¶
Maximal empirical accuracy (noise ceiling)¶
Maximal empirical accuracy is the upper bound for model performance based on the human choice variance within the same face triplets (often called the noise ceiling).
In the following we report the accuracy adjusted for the maximal empirical accuracy, followed by the bare accuracy in brackets.
Out[12]:
| sample_type | max_acc | min_n_samples | |
|---|---|---|---|
| session | |||
| 2D | multi-sub-sample | 0.6634 | 5.0 |
| 2D | multi-sub-sample_female | 0.6483 | 5.0 |
| 2D | multi-sub-sample_male | 0.6433 | 5.0 |
| 3D | multi-sub-sample | 0.6867 | 5.0 |
| 3D | multi-sub-sample_female | 0.6767 | 5.0 |
| 3D | multi-sub-sample_male | 0.7100 | 5.0 |
Sparse encoding models¶
SPoSE¶
2D Session: Accuracy of best performing SPoSE model 2D: 89.32% (59.26%) 3D Session: Accuracy of best performing SPoSE model 3D: 86.01% (59.06%)
VICE¶
2D Session: Accuracy of best performing VICE model 2D: 89.34% (59.27%) 3D Session: Accuracy of best performing VICE model 3D: 86.32% (59.27%)
2D Session: Accuracy of best performing VGGFace model '2023-11-15_04-44_VGGFaceHumanjudgmentFrozenCore' 2D: 85.92% (57.00%) 3D Session: Accuracy of best performing VGGFace model '2023-12-11_19-46_VGGFaceHumanjudgmentFrozenCore' 3D: 84.17% (57.80%)
Maximal emperical accuracy across viewing conditions is: 66.05%.
Now we take the same computation but across viewing conditions of the full-sample (i.e., data from the main study).
Predicting human choices in one viewing condition with the choices in the other condition
leads to an accuracy of 72.76% (48.06%).
Accuracy when gender matches¶
Calculate the probabilty when participants would only choose the odd-one-out based on gender. This would be a theoretical upper bound for an accuracy based on such a simply choice-behavior.
That is, when there is a triplet with, e.g., M-M-F, they choose F. For triplets with one gender only, they guess randomly.
The theoretical highest accuracy when predicting only mixed gender trials correctly, would be at: 75.76% The theoretical highest accuracy when predicting mixed gender trials correctly and one-gender trials on chance level, would be at: 83.84%
For now, this is just of theoretical interest, and might be reported in the appendix of the main paper.
Predictions within gender trials¶
Maximial empirical accuracy within gender-exclusive-trials¶
Out[22]:
| sample_type | max_acc | min_n_samples | |
|---|---|---|---|
| session | |||
| 2D | multi-sub-sample_female | 0.6483 | 5.0 |
| 2D | multi-sub-sample_male | 0.6433 | 5.0 |
| 3D | multi-sub-sample_female | 0.6767 | 5.0 |
| 3D | multi-sub-sample_male | 0.7100 | 5.0 |
Sparse encoding models¶
SPoSE¶
2D Session: female: Accuracy of best performing SPoSE model 2D within exclusive 'female' trials: 87.75% (56.89%) male: Accuracy of best performing SPoSE model 2D within exclusive 'male' trials: 90.88% (58.46%) 3D Session: female: Accuracy of best performing SPoSE model 3D within exclusive 'female' trials: 85.43% (57.81%) male: Accuracy of best performing SPoSE model 3D within exclusive 'male' trials: 81.30% (57.72%)
VICE¶
2D Session: female: Accuracy of best performing VICE model 2D within exclusive 'female' trials: 87.52% (56.74%) male: Accuracy of best performing VICE model 2D within exclusive 'male' trials: 89.80% (57.77%) 3D Session: female: Accuracy of best performing VICE model 3D within exclusive 'female' trials: 86.01% (58.20%) male: Accuracy of best performing VICE model 3D within exclusive 'male' trials: 81.59% (57.93%)
2D Session: female: Accuracy of best performing VGGFace model '2023-11-24_14-57_VGGFaceHumanjudgmentFrozenCore' 2D within exclusive 'female' trials: 82.52% (53.50%) male: Accuracy of best performing VGGFace model '2023-11-26_09-05_VGGFaceHumanjudgmentFrozenCore' 2D within exclusive 'male' trials: 85.34% (54.90%) 3D Session: female: Accuracy of best performing VGGFace model '2023-11-25_10-17_VGGFaceHumanjudgmentFrozenCore' 3D within exclusive 'female' trials: 79.21% (53.60%) male: Accuracy of best performing VGGFace model '2023-11-26_22-57_VGGFaceHumanjudgmentFrozenCore' 3D within exclusive 'male' trials: 78.45% (55.70%)
Representational differences between viewing conditions (2D & 3D)¶
Differences between the static 2D and the dynamic 3D condition
In [27]:
Copied!
rsa_corr_df = get_corr_df_rsm(corr_name=CORR_NAME, metric=METRIC)
# get_mer_table() # noqa: ERA001
rsa_corr_df = get_corr_df_rsm(corr_name=CORR_NAME, metric=METRIC)
# get_mer_table() # noqa: ERA001
In [28]:
Copied!
# Set N highest correlating VGG layers
n_highest_vgg: int = 5
# Set N highest correlating VGG layers
n_highest_vgg: int = 5
2D_BSM
Spearman correlation 2D_BSM ~ 3D_BSM_r: r=0.933, p<=0
Spearman correlation 2D_BSM ~ CFD_PFF_r: r=0.258, p<=5.2e-76
Spearman correlation 2D_BSM ~ DECA_SHAPE_r: r=0.259, p<=1e-76
Spearman correlation 2D_BSM ~ DECA_EXP_r: r=0.305, p<=2.3e-107
Spearman correlation 2D_BSM ~ DECA_POSE_r: r=0.089, p<=3.1e-10
Spearman correlation 2D_BSM ~ DECA_CAM_r: r=0.030, p<=0.038
Spearman correlation 2D_BSM ~ DECA_TEX_r: r=0.468, p<=2.2e-267
Spearman correlation 2D_BSM ~ DECA_DETAIL_r: r=0.319, p<=1.9e-117
Spearman correlation 2D_BSM ~ SPoSE_2D_r: r=0.927, p<=0
Spearman correlation 2D_BSM ~ VICE_2D_r: r=0.935, p<=0
Spearman correlation 2D_BSM ~ SPoSE_3D_r: r=0.914, p<=0
Spearman correlation 2D_BSM ~ VICE_3D_r: r=0.916, p<=0
Spearman correlation 2D_BSM ~ 2023-11-15_04-44_VGGFaceHumanjudgmentFrozenCore_2D_embedding_r: r=0.776, p<=0
Spearman correlation 2D_BSM ~ 2023-11-15_04-44_VGGFaceHumanjudgmentFrozenCore_2D_decision_r: r=0.876, p<=0
Correlation VGG layers
Spearman correlation 2D_BSM ~ VGG_3D-recon_MAXP_5_3_r: r=0.489, p<=4.8e-296
Spearman correlation 2D_BSM ~ VGG_3D-recon_RELU_5_3_r: r=0.482, p<=2.4e-286
Spearman correlation 2D_BSM ~ VGG_3D-recon_CONV_5_3_r: r=0.482, p<=2.4e-286
Spearman correlation 2D_BSM ~ VGG_3D-recon_FC6_r: r=0.432, p<=1.8e-224
Spearman correlation 2D_BSM ~ VGG_3D-recon_FC6-DROPOUT_r: r=0.432, p<=1.8e-224
******************************************************************************** 3D_BSM Spearman correlation 3D_BSM ~ 2D_BSM_r: r=0.933, p<=0 Spearman correlation 3D_BSM ~ CFD_PFF_r: r=0.246, p<=2.2e-69 Spearman correlation 3D_BSM ~ DECA_SHAPE_r: r=0.249, p<=4.7e-71 Spearman correlation 3D_BSM ~ DECA_EXP_r: r=0.299, p<=5.9e-103 Spearman correlation 3D_BSM ~ DECA_POSE_r: r=0.093, p<=6.9e-11 Spearman correlation 3D_BSM ~ DECA_CAM_r: r=0.041, p<=0.0042 Spearman correlation 3D_BSM ~ DECA_TEX_r: r=0.478, p<=4.8e-281 Spearman correlation 3D_BSM ~ DECA_DETAIL_r: r=0.326, p<=1.4e-122 Spearman correlation 3D_BSM ~ SPoSE_2D_r: r=0.912, p<=0 Spearman correlation 3D_BSM ~ VICE_2D_r: r=0.916, p<=0 Spearman correlation 3D_BSM ~ SPoSE_3D_r: r=0.936, p<=0 Spearman correlation 3D_BSM ~ VICE_3D_r: r=0.939, p<=0 Spearman correlation 3D_BSM ~ 2023-12-11_19-46_VGGFaceHumanjudgmentFrozenCore_3D_embedding_r: r=0.721, p<=0 Spearman correlation 3D_BSM ~ 2023-12-11_19-46_VGGFaceHumanjudgmentFrozenCore_3D_decision_r: r=0.879, p<=0
Correlation VGG layers
Spearman correlation 3D_BSM ~ VGG_3D-recon_MAXP_5_3_r: r=0.490, p<=5.4e-297
Spearman correlation 3D_BSM ~ VGG_3D-recon_RELU_5_3_r: r=0.483, p<=3e-287
Spearman correlation 3D_BSM ~ VGG_3D-recon_CONV_5_3_r: r=0.483, p<=3e-287
Spearman correlation 3D_BSM ~ VGG_3D-recon_FC6_r: r=0.435, p<=4.8e-228
Spearman correlation 3D_BSM ~ VGG_3D-recon_FC6-DROPOUT_r: r=0.435, p<=4.8e-228
********************************************************************************
Correlations within exclusive gender trials (RSA)¶
2D_BSM_female_only
Spearman correlation 2D_BSM_female_only ~ 3D_BSM_female_only_r: r=0.839, p<=0
Spearman correlation 2D_BSM_female_only ~ 3D_BSM_male_only_r: r=0.036, p<=0.21
Spearman correlation 2D_BSM_female_only ~ 2D_BSM_male_only_r: r=0.046, p<=0.11
Spearman correlation 2D_BSM_female_only ~ CFD_PFF_female_only_r: r=0.216, p<=2e-14
Spearman correlation 2D_BSM_female_only ~ DECA_SHAPE_female_only_r: r=0.155, p<=4.8e-08
Spearman correlation 2D_BSM_female_only ~ DECA_EXP_female_only_r: r=0.150, p<=1.3e-07
Spearman correlation 2D_BSM_female_only ~ DECA_POSE_female_only_r: r=0.109, p<=0.00013
Spearman correlation 2D_BSM_female_only ~ DECA_CAM_female_only_r: r=0.095, p<=0.00085
Spearman correlation 2D_BSM_female_only ~ DECA_TEX_female_only_r: r=0.154, p<=5.6e-08
Spearman correlation 2D_BSM_female_only ~ DECA_DETAIL_female_only_r: r=0.064, p<=0.024
Spearman correlation 2D_BSM_female_only ~ SPoSE_2D_female_only_r: r=0.923, p<=0
Spearman correlation 2D_BSM_female_only ~ VICE_2D_female_only_r: r=0.873, p<=0
Spearman correlation 2D_BSM_female_only ~ SPoSE_3D_female_only_r: r=0.814, p<=2.2e-291
Spearman correlation 2D_BSM_female_only ~ VICE_3D_female_only_r: r=0.824, p<=6.4e-304
Spearman correlation 2D_BSM_female_only ~ 2023-11-24_14-57_VGGFaceHumanjudgmentFrozenCore_2D_female_only_r: r=0.543, p<=6.8e-95
Spearman correlation 2D_BSM_female_only ~ 2023-11-15_04-44_VGGFaceHumanjudgmentFrozenCore_2D_decision_female_only_r: r=0.741, p<=7.3e-214
Correlation VGG layers
Spearman correlation 2D_BSM_female_only ~ VGG_3D-recon_RELU_5_3_female_only_r: r=0.436, p<=6.3e-58
Spearman correlation 2D_BSM_female_only ~ VGG_3D-recon_CONV_5_3_female_only_r: r=0.436, p<=6.3e-58
Spearman correlation 2D_BSM_female_only ~ VGG_3D-recon_MAXP_5_3_female_only_r: r=0.431, p<=1.9e-56
Spearman correlation 2D_BSM_female_only ~ VGG_3D-recon_RELU_5_2_female_only_r: r=0.398, p<=7.4e-48
Spearman correlation 2D_BSM_female_only ~ VGG_3D-recon_CONV_5_2_female_only_r: r=0.398, p<=7.4e-48
-------------------------------------------------------------------------------- 2D_BSM_male_only Spearman correlation 2D_BSM_male_only ~ 3D_BSM_female_only_r: r=0.055, p<=0.057 Spearman correlation 2D_BSM_male_only ~ 2D_BSM_female_only_r: r=0.046, p<=0.11 Spearman correlation 2D_BSM_male_only ~ 3D_BSM_male_only_r: r=0.881, p<=0 Spearman correlation 2D_BSM_male_only ~ CFD_PFF_male_only_r: r=0.196, p<=5e-12 Spearman correlation 2D_BSM_male_only ~ DECA_SHAPE_male_only_r: r=0.228, p<=5.8e-16 Spearman correlation 2D_BSM_male_only ~ DECA_EXP_male_only_r: r=0.245, p<=3.6e-18 Spearman correlation 2D_BSM_male_only ~ DECA_POSE_male_only_r: r=0.064, p<=0.026 Spearman correlation 2D_BSM_male_only ~ DECA_CAM_male_only_r: r=-0.004, p<=0.9 Spearman correlation 2D_BSM_male_only ~ DECA_TEX_male_only_r: r=0.255, p<=1.2e-19 Spearman correlation 2D_BSM_male_only ~ DECA_DETAIL_male_only_r: r=0.126, p<=1e-05 Spearman correlation 2D_BSM_male_only ~ SPoSE_2D_male_only_r: r=0.917, p<=0 Spearman correlation 2D_BSM_male_only ~ VICE_2D_male_only_r: r=0.877, p<=0 Spearman correlation 2D_BSM_male_only ~ SPoSE_3D_male_only_r: r=0.863, p<=0 Spearman correlation 2D_BSM_male_only ~ VICE_3D_male_only_r: r=0.834, p<=1.6e-318 Spearman correlation 2D_BSM_male_only ~ 2023-11-26_09-05_VGGFaceHumanjudgmentFrozenCore_2D_male_only_r: r=0.715, p<=2.3e-192 Spearman correlation 2D_BSM_male_only ~ 2023-11-15_04-44_VGGFaceHumanjudgmentFrozenCore_2D_decision_male_only_r: r=0.821, p<=1.1e-299
Correlation VGG layers
Spearman correlation 2D_BSM_male_only ~ VGG_3D-recon_FC6-DROPOUT_male_only_r: r=0.445, p<=1.1e-60
Spearman correlation 2D_BSM_male_only ~ VGG_3D-recon_FC6-RELU_male_only_r: r=0.445, p<=1.1e-60
Spearman correlation 2D_BSM_male_only ~ VGG_3D-recon_FC6_male_only_r: r=0.445, p<=1.1e-60
Spearman correlation 2D_BSM_male_only ~ VGG_3D-recon_MAXP_5_3_male_only_r: r=0.441, p<=1.7e-59
Spearman correlation 2D_BSM_male_only ~ VGG_3D-recon_RELU_5_3_male_only_r: r=0.432, p<=9.2e-57
-------------------------------------------------------------------------------- 3D_BSM_female_only Spearman correlation 3D_BSM_female_only ~ 2D_BSM_female_only_r: r=0.839, p<=0 Spearman correlation 3D_BSM_female_only ~ 3D_BSM_male_only_r: r=0.045, p<=0.11 Spearman correlation 3D_BSM_female_only ~ 2D_BSM_male_only_r: r=0.055, p<=0.057 Spearman correlation 3D_BSM_female_only ~ CFD_PFF_female_only_r: r=0.222, p<=4e-15 Spearman correlation 3D_BSM_female_only ~ DECA_SHAPE_female_only_r: r=0.128, p<=7e-06 Spearman correlation 3D_BSM_female_only ~ DECA_EXP_female_only_r: r=0.115, p<=5.9e-05 Spearman correlation 3D_BSM_female_only ~ DECA_POSE_female_only_r: r=0.112, p<=8.4e-05 Spearman correlation 3D_BSM_female_only ~ DECA_CAM_female_only_r: r=0.104, p<=0.00028 Spearman correlation 3D_BSM_female_only ~ DECA_TEX_female_only_r: r=0.169, p<=2.6e-09 Spearman correlation 3D_BSM_female_only ~ DECA_DETAIL_female_only_r: r=0.074, p<=0.01 Spearman correlation 3D_BSM_female_only ~ SPoSE_2D_female_only_r: r=0.827, p<=2.4e-308 Spearman correlation 3D_BSM_female_only ~ VICE_2D_female_only_r: r=0.820, p<=5.4e-298 Spearman correlation 3D_BSM_female_only ~ SPoSE_3D_female_only_r: r=0.928, p<=0 Spearman correlation 3D_BSM_female_only ~ VICE_3D_female_only_r: r=0.886, p<=0 Spearman correlation 3D_BSM_female_only ~ 2023-11-25_10-17_VGGFaceHumanjudgmentFrozenCore_3D_female_only_r: r=0.630, p<=1.6e-136 Spearman correlation 3D_BSM_female_only ~ 2023-12-11_19-46_VGGFaceHumanjudgmentFrozenCore_3D_decision_female_only_r: r=0.796, p<=2.7e-269
Correlation VGG layers
Spearman correlation 3D_BSM_female_only ~ VGG_3D-recon_RELU_5_3_female_only_r: r=0.410, p<=8e-51
Spearman correlation 3D_BSM_female_only ~ VGG_3D-recon_CONV_5_3_female_only_r: r=0.410, p<=8e-51
Spearman correlation 3D_BSM_female_only ~ VGG_3D-recon_MAXP_5_3_female_only_r: r=0.404, p<=2.1e-49
Spearman correlation 3D_BSM_female_only ~ VGG_3D-recon_RELU_5_2_female_only_r: r=0.377, p<=1.1e-42
Spearman correlation 3D_BSM_female_only ~ VGG_3D-recon_CONV_5_2_female_only_r: r=0.377, p<=1.1e-42
-------------------------------------------------------------------------------- 3D_BSM_male_only Spearman correlation 3D_BSM_male_only ~ 3D_BSM_female_only_r: r=0.045, p<=0.11 Spearman correlation 3D_BSM_male_only ~ 2D_BSM_female_only_r: r=0.036, p<=0.21 Spearman correlation 3D_BSM_male_only ~ 2D_BSM_male_only_r: r=0.881, p<=0 Spearman correlation 3D_BSM_male_only ~ CFD_PFF_male_only_r: r=0.176, p<=5e-10 Spearman correlation 3D_BSM_male_only ~ DECA_SHAPE_male_only_r: r=0.202, p<=9e-13 Spearman correlation 3D_BSM_male_only ~ DECA_EXP_male_only_r: r=0.242, p<=9.1e-18 Spearman correlation 3D_BSM_male_only ~ DECA_POSE_male_only_r: r=0.069, p<=0.016 Spearman correlation 3D_BSM_male_only ~ DECA_CAM_male_only_r: r=0.014, p<=0.63 Spearman correlation 3D_BSM_male_only ~ DECA_TEX_male_only_r: r=0.276, p<=6.4e-23 Spearman correlation 3D_BSM_male_only ~ DECA_DETAIL_male_only_r: r=0.132, p<=3.7e-06 Spearman correlation 3D_BSM_male_only ~ SPoSE_2D_male_only_r: r=0.846, p<=0 Spearman correlation 3D_BSM_male_only ~ VICE_2D_male_only_r: r=0.838, p<=0 Spearman correlation 3D_BSM_male_only ~ SPoSE_3D_male_only_r: r=0.945, p<=0 Spearman correlation 3D_BSM_male_only ~ VICE_3D_male_only_r: r=0.883, p<=0 Spearman correlation 3D_BSM_male_only ~ 2023-11-26_22-57_VGGFaceHumanjudgmentFrozenCore_3D_male_only_r: r=0.671, p<=7.7e-161 Spearman correlation 3D_BSM_male_only ~ 2023-12-11_19-46_VGGFaceHumanjudgmentFrozenCore_3D_decision_male_only_r: r=0.832, p<=2.5e-315
Correlation VGG layers
Spearman correlation 3D_BSM_male_only ~ VGG_3D-recon_FC6_male_only_r: r=0.449, p<=6.9e-62
Spearman correlation 3D_BSM_male_only ~ VGG_3D-recon_FC6-DROPOUT_male_only_r: r=0.449, p<=6.9e-62
Spearman correlation 3D_BSM_male_only ~ VGG_3D-recon_FC6-RELU_male_only_r: r=0.449, p<=6.9e-62
Spearman correlation 3D_BSM_male_only ~ VGG_3D-recon_MAXP_5_3_male_only_r: r=0.441, p<=1.8e-59
Spearman correlation 3D_BSM_male_only ~ VGG_3D-recon_RELU_5_3_male_only_r: r=0.433, p<=3.4e-57
--------------------------------------------------------------------------------
Behaviorally relevant face features¶
Correlations of VICE weights (dimensions) with PFA (CFD).
Also computed in computational_choice_model.py; and, see plots above.
After VICE optimization: * in the 2D-session, 28 dimensions remained relevant. * in the 3D-session, 26 dimensions remained relevant.
3 strongest correlations between first 3 VICE dimensions and PFAs 2D-VICE dimension 1: PFA labels MidcheekChinR MidcheekChinL CheeksAvg R 0.31 0.31 0.31 2D-VICE dimension 2: PFA labels EyeSize CheeksAvg MidcheekChinL R -0.46 0.48 0.48 2D-VICE dimension 3: PFA labels NoseWidth BottomLipChin NoseShape R -0.47 -0.48 -0.58 *-*-*-*-*-*-*-*-*-*-*-*-*-*-*-*-*-*-*-*-*-*-*-*-*-*-*-*-*-*-*-*-*-*-*-*-*-*-*-*-*-*-*-*-*-*-*- 3D-VICE dimension 1: PFA labels CheeksAvg MidcheekChinL BottomLipChin R 0.25 0.26 0.3 3D-VICE dimension 2: PFA labels UpperFaceLength2 fWHR2 NoseShape R 0.57 -0.6 -0.61 3D-VICE dimension 3: PFA labels EyeSize ChinLength BottomLipChin R 0.35 -0.41 -0.48 *-*-*-*-*-*-*-*-*-*-*-*-*-*-*-*-*-*-*-*-*-*-*-*-*-*-*-*-*-*-*-*-*-*-*-*-*-*-*-*-*-*-*-*-*-*-*-
Female eye-size: count 50.000000 mean 0.054131 std 0.006044 min 0.039763 25% 0.051509 50% 0.052956 75% 0.057007 max 0.071462 Name: P056, dtype: float64 Male eye-size: count 50.000000 mean 0.048736 std 0.007077 min 0.029462 25% 0.043373 50% 0.049134 75% 0.053614 max 0.061582 Name: P056, dtype: float64 ttest_ind(a=female_feat_tabel[eye_size_code], b=male_feat_tabel[eye_size_code]) = Ttest_indResult(statistic=4.099633006037346, pvalue=8.543611806806524e-05)