SE-AlexNet

Facial emotion recognition with channel recalibration

We tested how SE reduction ratio, insertion position, and pretraining affect emotion responses. Masked faces and Grad-CAM then let us examine which facial regions contribute to those responses.

Code & stimuli ↗GitHub

Model files ↗Hugging Face

Reduction ratio and emotion response

Experiment 1 · Face and object pretraining at R = 2, 4, 8, 16, and 32

The SE gate rescales a feature vector before classification.

AlexNet FC7→SE gate→FC8

Feature recalibration Equation 1: feature recalibration

(1)

Reduction ratio Equation 2: output width equals input width divided by R

(2)

Reported result. The face-pretrained, late-insertion model at R = 16 has PSE 0.51 and RMSE 0.09 in the full-face condition; mean human PSE is approximately 0.52.

PSE comparison in the final manuscript — PSE comparison from the final manuscript. Open full figure ↗

Psychometric curves for the face-pretrained late-insertion models — Psychometric response curves · full-face condition. Fitting code and data ↗

Behavioral analysis equations (3–5)

Psychometric curve Equation 3: psychometric response function

(3)

Curve error Equation 4: root mean square model–human curve error

(4)

PSE difference Equation 5: absolute difference between model and human PSE

(5)

SE insertion position

Experiment 2 · Early, middle, and late insertion in the classifier pathway

Late insertion places SE between FC7 and FC8.

Final-manuscript comparison of SE insertion positions — Insertion-position comparison from the final manuscript. Open full figure ↗

Occlusion and facial-region attention

Experiment 3 · Four masking conditions, Grad-CAM, and region analysis

Choose a pretraining source, insertion position, and reduction ratio. The four masking conditions update together using the representative panels in the manuscript’s Grad-CAM figure.

Pretraining

Model

Reduction

The unoverlaid inputs are aligned stimulus 11, included to explain each masking condition. The heatmaps are representative panels from final-manuscript Figure 7; the manuscript does not identify them as stimulus 11 or a same-image model comparison. View complete Figure 7 ↗

Region-level readout

Grad-CAM saliency was summarized in fixed eye, nose, and mouth regions on aligned 224 × 224 stimuli. Coordinates are image pixels.

Aligned face showing the predefined eye, nose, and mouth regions — ROI boundaries on an aligned face

Region	x	y
Eyes	52–171	70–99
Nose	52–171	120–149
Mouth	52–171	169–199

For full faces, the reported diagnostic ratio was 0.95 ± 0.03 for face pretraining and 0.77 ± 0.03 for object pretraining (mean ± SEM; p < 0.001). Eye-region saliency was 0.21 versus 0.07 (p = 0.006).

Grad-CAM and ROI equations (6–9)

Channel weight Equation 6: Grad-CAM channel weight