面向高效面孔认知的压缩激励计算模型
A squeeze-and-excitation computational model for efficient face cognition
彭玉佳 1包家瑜 2赵振华 3应浩江4
作者信息
- 1. 北京通用人工智能研究院跨媒体通用人工智能全国重点实验室, 北京 100080;北京大学人工智能研究院, 北京 100871;北京大学心理与认知科学学院、北京市行为与心理健康重点实验室、生物与机器智能教育部重点实验室, 北京 100871
- 2. 苏州大学教育学院心理学系, 苏州 215123;香港城市大学神经科学系, 香港 999077
- 3. 苏州大学教育学院心理学系, 苏州 215123;北京师范大学认知神经与学习国家重点实验室, 北京 100875;IDG/麦戈文脑科学研究院, 北京 100875
- 4. 苏州大学教育学院心理学系, 苏州 215123
- 折叠
摘要
人类面孔认知可能依赖于对高维视觉信息的压缩以及对诊断性特征的选择性增强。受灵长类面孔加工系统低维表征研究的启发，本研究在AlexNet中引入压缩激励（squeeze-and-excitation, SE）模块，考察压缩比、模块插入位置和预训练背景如何影响模型的面孔情绪判断及其与人类行为的对应程度。研究通过拟合心理物理函数获得主观相等点（PSE），并衡量模型反应曲线与人类平均反应曲线之间的差异。实验1比较了五种压缩比；实验2比较了SE模块在全连接层前序与后序位置的表现；实验3结合梯度加权类别激活映射（Grad-CAM）和局部遮挡，分析模型决策对不同面孔区域的敏感性。结果表明，在所考察的模型配置中，采用面孔预训练权重、压缩比为16并将SE模块插入FC7与FC8之间的模型，与人类行为表现出最小的总体差异。该优势取决于压缩比、插入位置和预训练背景的共同作用。Grad-CAM和遮挡结果进一步表明，模型决策对眼部、嘴部等面孔区域较为敏感。上述结果说明，在特定架构与训练条件下，SE特征重标定能够改善卷积神经网络输出与人类面孔情绪判断之间的对应程度，并为利用压缩表征研究面孔认知提供了一种可检验的计算建模方案。
Abstract
Efficient facial emotion perception may depend on compressing high-dimensional visual information while selectively preserving diagnostically relevant features (e.g., the mouth, eyes). Inspired by evidence for low-dimensional representations in primate face-processing systems, we incorporated squeeze-and-excitation (SE) modules into AlexNet and examined how the compression ratio, module position, and pretraining background affected facial-emotion judgments and their correspondence with human behavior. Psychometric functions were fitted to obtain points of subjective equality (PSEs), and to quantify the discrepancy between each models response curve and the mean human response curve. Experiment 1 compared five SE compression ratios. Experiment 2 compared SE modules inserted at earlier and later positions in the fully connected layers. Experiment 3 combined gradient-weighted class activation mapping (Grad-CAM) method with facial feature occlusion to characterize the models sensitivity to different facial regions. Among the configurations examined, the face-pretrained model with a compression ratio of 16 and an SE module inserted between FC7 and FC8 showed the smallest overall discrepancy from human behavior. This advantage depended jointly on compression ratio, insertion position, and pretraining background. The Grad-CAM and occlusion results further indicated that model decisions were sensitive to facial regions including the eyes and mouth. Together, these findings show that, under specific architectural and training conditions, SE-based feature recalibration can improve the correspondence between convolutional neural network outputs and human facial-emotion judgments. The study therefore provides a testable computational approach for investigating compressive representations in human face perception.关键词
表征压缩/卷积神经网络/视知觉/面孔认知/认知机制Key words
representational compression/convolutional neural network/visual perception/face cognition/cognitive mechanism引用本文复制引用
彭玉佳,包家瑜,赵振华,应浩江.面向高效面孔认知的压缩激励计算模型[EB/OL].(2026-10-18)[2026-10-22].https://chinaxiv.org/abs/202610.00102.学科分类
计算技术、计算机技术
