Neural and computational mechanisms underlying robust object recognition

点击翻译按钮获取中文摘要

Deep neural networks (DNNs) for object classification have been argued to provide the most promising state- of-the-art models of the visual system, accompanied by claims that they have attained or even surpassed human-level performance. However, mounting evidence has revealed that DNNs fail catastrophically when faced with more noisy or degraded viewing conditions. By contrast, the human visual system is far more robust. To better understand and model human vision, one must determine whether the brittle nature of DNN performance arises from flaws in their architectural design, imperfections in their learning protocols, or inadequate sampling of relevant training experiences. This project will investigate the neurocomputational bases of robust object recognition, focusing on challenge conditions of visual noise and blur, to develop new DNN models that can provide a better account of human behavioral and neural responses to object images that will vary from clear to severely degraded. Both feedforward and recurrent DNN architectures will be evaluated, and the critical sets of training experiences needed for DNNs to attain robustness will be determined. In Aim 1, we will evaluate what types of DNNs can adequately predict human behavioral and neural responses to objects embedded in noise on an image-by-image basis. Correspondences between fMRI responses at multiple levels of the human visual pathway will be compared with layer-wise DNN representations to evaluate the goodness of fit for DNN model predictions. In Aim 2, we will determine what types of DNNs can better account for human behavioral and neural responses to blurry object images. We will further explore how training with blurry images modifies the visual representations learned by DNNs, leading to greater robustness to other types of image degradation and greater sensitivity to shape information. In Aim 3, we will investigate whether perceptual training with noisy or blurry objects can allow humans to acquire even greater robustness. We will then determine whether human improvements in behavioral and neural performance can be effectively modeled by DNNs that undergo comparable regimens in visual training. As a whole, this project will lead to the development of powerful new DNN models that provide a better account of human behavioral and neural responses across a wide range of challenging viewing conditions. By developing a better neurocomputational model of the intact human visual system, we will be better positioned to eventually develop models of central visual disorders, which can arise from neurodevelopmental or neurological disorders, stroke, head injury, brain tumors or other diseases. The advancement of more robust, human-like DNNs is also highly relevant to AI applications in computer vision and medical image processing.