SE-AlexNet

Squeeze-and-Excitation Algorithm

for Efficient Face Recognition

GitHub Dataset Weights

Author List: author anonymous were required druing peer review process

SE-AlexNet Grad-CAM heatmap with eyes and mouth attention

Input stimulus SE-AlexNet attention

5 architectures

5 reduction ratios

4 stimulus conditions

21 emotion-intensity samples

Question

Does channel-wise reduction change how CNNs use facial regions?

Visual evidence

Aligned stimuli make Grad-CAM differences directly comparable.

Quantification

ROI saliency and fitting curves connect attention with behavior.

GradCAM Result

Explicit comparison between all conditions

Same face, different models, directly comparable attention.

Full face stimulus — A 224 by 224 aligned full-face stimulus used for model probing

Start with the same face

The aligned input makes attention the visual variable

SE blocks

Turn channels into a controlled bottleneck

Channel descriptors are squeezed, excited, and used to reweight feature maps

x ∈ R^B×C×H×W

z_c = (H·W)^-1Σ_i,jx_c(i,j)

s = σ(W₂ δ(W₁z)), W₁: C → C/r

x' = x ⊙ s

AlexNet feature maps SE channel gate Classifier readout

Insertion test SE blocks are inserted at candidate AlexNet locations

Reduction test r controls the channel bottleneck: 2, 4, 8, 16, and 32

Visual test Grad-CAM checks whether attention shifts on aligned faces

Attention Region

Measured where faces carry emotion cues

Grad-CAM is summarized inside eyes, nose, and mouth bands on aligned 224 x 224 faces.

Eyes Nose Mouth

Proportion of saliency

A region score is the share of total Grad-CAM activation inside that ROI.

PoS_ROI = Σ_(x,y)∈ROI L^c_Grad-CAM(x,y) Σ_x,y L^c_Grad-CAM(x,y)

Eyes: x 52-171, y 70-99
Nose: x 52-171, y 120-149
Mouth: x 52-171, y 169-199

Vertical Trend

X-Axis: Accumulated Activation (Saliency Strength).

Y-Axis: Pixel Height (0 = Top of Face, 224 = Bottom of Face).

The black curve represents the distribution of attention from the top to the bottom of the face.

Gaussian Smoothing (Sigma=3) is applied to reduce noise.

If the black curve's peak aligns with the Red Band, the model is primarily focusing on the Mouth.

Results

Core conclusion

Late SE insertion changes the spatial evidence used by AlexNet while preserving a measurable behavioral readout.

Analysis 1

Psychometric fitting

Curve shape summarizes model behavior across emotion intensity.

Reproduce this analysis in our PsychometricFittingCurve repository.

Psychometric fitting curve for Location-3 FaceBased full-face condition — Late SE-AlexNet, FaceBased, full-face condition.

Psychometric fitting curve for SE-FC L3 FaceBased full-face condition — Additional SE insertion setting for comparison.

Analysis 2

Training dynamics

Accuracy and loss check that the visual effect is not detached from optimization.

Training accuracy curve — Accuracy, Location-3 FaceBased R32.

Training loss curve — Loss, same training condition.

Analysis 3

ROI and attention distribution

ROI proportions are the main quantitative readout; trend plots show where the vertical peak lands.

Reproduce Grad-CAM and ROI saliency in our GradCAM-ROI-SaliencyMapDecoder repository.

Average ROI saliency by condition — Mean ROI saliency by masked-face condition.

Readout ROI scores turn Grad-CAM from a heatmap into region-level evidence.

Alignment The same ROI bands make model-to-model attention shifts comparable.

Which region carries the strongest share?

How to cite

Please use the manuscript citation for this project.

@article{,
              title={},
              author={},
              year={},
              note={Manuscript under review}
            }

Resources

Open dataset & model weights

Code, Dataset and Model weights are publicly available on Github and Hugging Face for research use. Please cite the project and use the released files for evaluation and reproducibility.

GitHub Dataset Weights