Closing the Gap Between High- and Low-Dimensional Models of High-Level Vision

我们的视觉感知是高度复杂的，以捕捉我们周围视觉世界的复杂性。因此，最近的视觉感知模型，如深度神经网络模型，变得非常复杂，也能成功地捕捉人类对任意照片的感知。然而，这些模型还没有将对象表示为单独的实体，它们使用了为其他目的而设计的特征，如对象识别或纹理生成。此外，它们的高维性造成了统计问题。这些方面在这些处理自然刺激的高维模型和我们所理解的低维、基于对象的模型之间造成了差距。在这个项目中，我的目标是通过双方的合作来缩小这一差距。从高维模型的角度出发，我的第一个目标是改进我们比较高维模型和更简单模型的方法。在这里，我想介绍现代基于稀疏性的统计方法，这些方法是专门为高维问题设计的。此外，我希望改进表征相似性分析——高维模型和数据的少数现有方法之一。从更简单的、可解释的模型开始，我想通过将可解释的模型推广到由随机叠加的简单形状组成的枯叶刺激来研究感知组织。这些刺激很有趣，因为它们是分析上易于处理的图像刺激，尽管如此，它们还是孤立了感知组织问题。我计划用这些刺激在人类身上做一个行为实验来测试我们感知组织的能力。对这些结果进行建模将产生广泛的模型来应用目标1中的方法。最后，我的第三个目标是利用我从前两个目标中获得的见解来改进我们对自然刺激的高级感知模型。我将使用我第一个目标中改进的统计方法来修剪和重组现有的深度学习模型，以提高它们作为人类视觉感知模型的能力。此外，我将进一步将我从枯叶刺激中获得的关于物体、遮挡和分割的见解推广到自然图像。因此，这个项目的结果可能有助于更好的人类感知的图像计算模型，它可以将视觉世界分成物体，其内部工作被明确地选择来模拟人类的处理。

Our visual perception is highly complex to capture the complexity of our surrounding visualworld. As a consequence of this, recent models of visual perception like deep neural network models became highly complex as well to capture human perception of arbitrary photographs with some success. However, these models do not yet represent objects as separate entities and they use features designed for other purposes like object recognition or texture generation. Furthermore, their high dimensionality creates statistical problems. These aspects create a gap between these high-dimensional models which work on natural stimuli and the low-dimensional, object-based models we understand. With this project I aim to close this gap by working from both sides. Starting from the side of high-dimensional models, my first aim is to improve our methods to compare high-dimensional models to simpler models. Here I want to introduce modern sparsity based statistical methods, which were specifically designed for high-dimensional problems. Furthermore, I hope to improve upon the representational similarity analysis—one of the few existing methods for high-dimensional models and data. Starting from the side of simpler, interpretable models, I want to work on perceptual organization by generalizing interpretable models to dead-leaves stimuli, which consist of randomly superimposed simple shapes. These stimuli are interesting, because they are analytically tractable image stimuli, which nonetheless isolate the perceptual organization problem. I plan to use these stimuli to perform a behavioural experiment in humans to test our capability for perceptual organization. Modelling these results will result in a broad range of models to apply the methods from aim 1. Finally, my third aim is to use the insights I gain from my first two aims to improve our models of high-level perception of natural stimuli. I am going to use the improved statistical methods from my first aim to prune and recombine existing deep learning models to improve their capabilities as models of human visual perception. Furthermore, I am going to further generalise insights I gain about objects, occlusion and segmentation from the dead-leaves stimuli to natural images. Thus, the results of this project may contribute to better image-computable models of human perception, which can separate the visual world into objects and whose internal workings are explicitly selected to model human processing.