RI: Small: Integrating Flexible Normalization Models of Visual Cortex into Deep Neural Networks

点击翻译按钮获取中文摘要

Recent advances in artificial intelligence models of deep neural networks have led to tremendous progress in artificial systems that recognize objects in scenes, and in a host of other applications such as speech recognition, and robotics. Although deep neural networks often incorporate computations inspired by the brain, these have typically been applied in a fairly simple and restrictive manner, rather than based on more principled models of neural processing in the brain. Using vision as a paradigmatic example, this project proposes that artificial systems can benefit from integrating approaches that have been developed in biological models of neural processing of scenes. The biological models make use of contextual flexibility, whereby neurons are influenced in a rich way by the image structure that spatially surrounds a given object or feature. This flexibility is expected to improve task performance in deep neural networks, and to impact development of artificial systems that are more compatible with human cognition. The resulting framework, with its deep architecture spanning multiple layers of processing, will, in turn, make predictions about neural processing in the brain, which will impact the neuroscience and cognitive science communities. This project focuses specifically on normalization, a nonlinear computation that is ubiquitous in the brain, and that has been shown to benefit task performance in deep neural networks. The project will develop more principled strategies for determining normalization in deep convolutional neural networks. The main focus will be on learning a form of flexible normalization based on scene statistics models of visual cortex. In this framework, normalization is recruited only to the degree that a visual input is inferred to contain statistical dependencies across space. Performance will be tested for classification and segmentation on large-scale image databases, and will also target tasks more suited to mid-level vision such as figure/ground judgment. This will result in better understanding of normalization nonlinearities in deep convolutional networks, and the implications of flexible normalization for task performance and generalization compared to other forms of normalization. Biologically, normalization is poorly understood beyond primary visual cortex. The models developed will help shed light on the equivalence of this inference for middle cortical areas, and make predictions about what image structure leads to recruitment of normalization. This project will also include launching of an interdisciplinary Deep Learning Discussion Group.