Convolutional neural networks trained with a developmental sequence of blurry to clear images reveal core differences between face and object processing

Abstract

Although convolutional neural networks (CNNs) provide a promising model for understanding human vision, most CNNs lack robustness to challenging viewing conditions such as image blur, whereas human vision is much more reliable. Might robustness to blur be attributable to vision during infancy, given that acuity is initially poor but improves considerably over the first several months of life? Here, we evaluated the potential consequences of such early experiences by training CNN models on face and object recognition tasks while gradually reducing the amount of blur applied to the training images. For CNNs trained on blurry to clear faces, we observed sustained robustness to blur, consistent with a recent report by Vogelsang and colleagues (2018). By contrast, CNNs trained with blurry to clear objects failed to retain robustness to blur. Further analyses revealed that the spatial frequency tuning of the two CNNs was profoundly different. The blurry to clear face-trained network successfully retained a preference for low spatial frequencies, whereas the blurry to clear object-trained CNN exhibited a progressive shift toward higher spatial frequencies. Our findings provide novel computational evidence showing how face recognition, unlike object recognition, allows for more holistic processing. Moreover, our results suggest that blurry vision during infancy is insufficient to account for the robustness of adult vision to blurry objects.

Do training with blurred images make convolutional neural networks closer to humans concerning object recognition performance and internal representations?

“…We evaluated the object-recognition performance of the blur-trained CNNs not only using low-pass filtered test images, but also for other types of images including bandpass filtered images and shape-texture-cue-conflict images [Geirhos et al, 2019] to see whether blur training affects global configurational processing in general. In agreement with the previous reports [Avberšek et al, 2021, Jang and Tong, 2021], our results show that mixed training on sharp and blurred images (B+S) is more effective in making the CNNs robust against a change in image blur in object recognition in comparison with the training simulating the human development by gradually improving image sharpness during training (B2S). The B+S training however is not sufficient to produce robust human-like object recognition based on global configurational features.…”

“…A recent study using object recognition [Avberšek et al, 2021] reports the effect of training schedule consistent with ours. The task difference may be related to the fact that the optimal discriminative features for object recognition are biased toward high frequencies while only low-frequency features are sufficient for good face classification accuracy [Jang and Tong, 2021].…”

“…First, we visualized the receptive fi eld of the fi rst convolutional layer (Fig. 21), as previous studies did [Vogelsang et al, 2018, Jang andTong, 2021]. B+S-Net shows a shift of the spatial frequency tuning to the lower frequency side compared to S-Net.…”

“…It has been suggested that the experience of blurred visual images might be functionally beneficial, enabling the visual system to use global configural structures in image recognition [Grand et al, 2001, Le Grand et al, 2004, Vogelsang et al, 2018. Several recent studies use machine learning of artificial neural networks to computationally test this hypothesis [Vogelsang et al, 2018, Katzhendler and Weinshall, 2019, Avberšek et al, 2021, Jang and Tong, 2021. Vogelsang trained the convolutional network (CNN) to recognize human faces.…”

Embodied vision for learning object representations

“…This can aid the learning of reflexes, like the acquisition of vergence behavior (alignment of the two eyes on the same point) [30] or the detection of binocular disparity (difference between the two eye images) [31]. This can also aid learning representations of faces robust to different resolutions [32], [33].…”