关键词: 表征压缩, 卷积神经网络, 视知觉, 面孔认知, 认知机制

Abstract: Convolutional neural networks (CNNs) are inspired by biological vision, yet their success usually depends on deeper architectures and large-scale data. By contrast, primate face processing appears to rely on compact, low-dimensional representations that reduce redundancy while preserving useful information. This discrepancy raises a theoretically important question for cognitive science and artificial intelligence: can representational compression improve artificial face cognition in a way that is both behaviorally human-like and computationally efficient? Drawing on efficient-coding principles and neural evidence on face representation, we tested whether a lightweight squeeze-and-excitation (SE) mechanism would enhance AlexNet. We predicted that an appropriately tuned compression process, especially at a late semantic stage, would produce the closest match to human facial-emotion judgments. Three experiments were conducted in this study. In Experiment 1, the SE modules with five compression ratios were inserted between FC7 and FC8 of the AlexNet, and model behavior was compared with that of the original AlexNet, VGG16, with human participants’ performance as the benchmark. Forty participants categorized a series of morphed faces ranging from sad to happy, and psychometric functions were used to estimate PSEs for humans and models. All CNN models were fine-tuned for facial emotion recognition through transfer learning, using both object-pretrained and face-pretrained weights. In Experiment 2, the insertion position of the module was manipulated by placing it before FC6, between FC6 and FC7, or between FC7 and FC8, while the same procedures were maintained. In Experiment 3, model interpretability was examined with Grad-CAM visualizations and feature-occlusion tests with the eyes, nose, and mouth occluded. Results from three experiments showed that SE-AlexNet with a compression ratio of 16 achieved behavioral performance most closely aligned with human performance in a facial emotion recognition task. This advantage did not arise merely from adding an SE module, but depended jointly on compression strength and insertion stage. The stage of module insertion was also found to be critical, with the best performance observed when the module was inserted at a late stage of higher-level visual representation, specifically between FC7 and FC8. Grad-CAM visualizations further indicated the best-performing late-stage SE-AlexNet preferentially attended to the eyes and mouth, which are diagnostically important regions in human facial emotion perception. When these regions were occluded, model performance declined clearly, suggesting that the optimized network relied on human-relevant facial features. The findings together suggest that biologically inspired representational compression can improve a classical convolutional neural network without requiring a substantially deeper or more complicated architecture. More importantly, the effectiveness of compression is constrained by where and how strongly it is implemented. Compression is most beneficial when it operates on late, higher-level stage, which is consistent with the view that face processing depends on compact coding in higher visual processing stages. The results therefore provides computational support for the idea that human face cognition is guided by compressed representations and selective attention to diagnostically informative features. It also offers a practical framework for designing lightweight, interpretable, and cognitively plausible neural networks for facial emotion recognition and related face-processing tasks.

Key words: presentational compression, convolutional neural network (CNN), visual perception, face recognition, cognitive mechanism