Deep learning models challenge the prevailing assumption that face-like effects for objects of expertise support domain-general mechanisms

The question of whether perceptual expertise is mediated by general-expert or domain-specific processing mechanisms has been debated for decades. Because humans are experts in face recognition, face-like neural and cognitive effects for objects of expertise were considered to support for the general-expertise hypothesis. Conversely, stronger effects for faces than objects of expertise were considered to support the domain-specific hypothesis. However, the effects of domain, experience, and level of categorization, are confounded in human studies, which may lead to erroneous inferences. To overcome these limitations, we used computational models of perceptual expertise and tested different domains (objects, faces, birds) and levels of categorization (basic, sub-ordinate, individual) in isolation, matched for amount of experience. Like humans, the models generated a larger inversion effect for faces than for objects. Importantly, a face-like inversion effect was found for individual-based categorization of non-faces (birds) but only in a network specialized for that domain. Thus, contrary to prevalent assumptions, face-like effects in objects of expertise may originate from domain-specific rather than domain-general processing mechanisms. More generally, we show how deep learning algorithms can be used to isolate the effects of factors that are inherently confounded in the natural environment of biological organisms.

Methods
Creation of verification tests for upright and inverted images is described in the article.

Each verification test is composed of 2 txt files:

Containing pairs of images belonging to the same class (both either upright or inverted, labeled "same")
Containing pairs of images belonging to different classes (both either upright or inverted, labeled "diff")

To calculate the AUROC, calculate the distances between the image pairs in the corresponding "same" and "diff" files.

Corresponding txt files are found in the same directory, with the prefix "same" or "diff" and the suffix "_{id}.txt". For example, corresponding txt files used for the verification tests are "same_3.txt" and "diff_3.txt".

Distance between image representations according to the different models could also be found in appropriately named csv files (for example, "260_inanimate.csv" representations from the network trained on 260 classes from ImageNet).

Trained neural networks are available as PyTorch weights, and trained according to the procedure described in the paper.

感知专长究竟是由通用专长加工机制还是领域特异性加工机制介导的，这一议题已争论数十年。由于人类本身是面孔识别领域的专家，针对各类专长对象所表现出的类面孔神经与认知效应，曾被视作支持通用专长假说的依据。反之，面孔相较于各类专长对象所呈现出的更强效应，则被认为是领域特异性假说的佐证。然而，在人类相关研究中，领域类型、经验水平与分类层级三者存在固有混淆，这可能导致推论出现偏差。为克服这些局限，我们采用感知专长的计算模型，在匹配经验总量的前提下，单独测试了不同领域（物体、面孔、鸟类）与不同分类层级（基本层级、下位层级、个体层级）的效应。与人类受试者的表现一致，模型在面孔任务中产生的倒置效应显著大于物体任务。值得注意的是，在非面孔领域（鸟类）的个体层级分类任务中，同样观察到了类面孔的倒置效应，但该现象仅出现在针对该领域专门训练的网络中。因此，与当前主流假设相悖，各类专长对象所表现出的类面孔效应，可能源自领域特异性而非领域通用的加工机制。更广泛而言，本研究展示了如何借助深度学习算法，分离出在生物有机体的自然环境中固有混淆的各类因素效应。

方法
正置与倒置图像的验证测试构建方法详见本文。
每个验证测试由2个文本文件构成：
1. 包含属于同一类别的图像对（两类图像均为正置或均为倒置，标记为"same"）；
2. 包含属于不同类别的图像对（两类图像均为正置或均为倒置，标记为"diff"）。

为计算接收者操作特征曲线下面积（AUROC），需计算对应"same"与"diff"文件中图像对之间的表征距离。

对应文本文件位于同一目录下，前缀为"same"或"diff"，后缀为"_{id}.txt"。例如，用于验证测试的对应文本文件为"same_3.txt"与"diff_3.txt"。

不同模型下的图像表征间距离也可在命名规范匹配的csv文件中获取（例如，基于ImageNet 260个类别训练的网络所对应的表征文件"260_inanimate.csv"）。

训练完成的神经网络可通过PyTorch权重文件获取，其训练流程详见本文所述方法。