SFB 1233: Robust Vision - Inference Principles and Neural Mechanisms

协作研究中心(CRC)1233汇聚了机器学习、计算机视觉和系统神经科学领域的领先研究人员，以揭示稳健视觉背后的计算原理。视觉对于动物和人类在大脑对世界的内部模型和外部环境之间实现可靠的对应起着关键作用。它依赖于对分布在整个图像上的亮度信号中隐含的有意义的模式的感觉和解释。理解促进这些模式的稳健视觉的原理和算法在理解生物视觉中起着基础性的作用。近年来，随着算法在越来越具挑战性的问题上得到测试，计算机视觉取得了长足的进步，包括自动驾驶等现实世界的应用。然而，人类的视觉显然仍在稳健性方面表现出色。合作研究中心(CRC)利用计算机视觉方面的成功经验，寻求促进我们对视觉在生物学和机器中的稳健性的理解。在计算研究人员和神经学家的密切合作中，我们的目标是揭示强健视觉的原理，并确定其在哺乳动物大脑中的神经基础。特别是，CRC侧重于视觉的神经生物学与当前机器视觉算法显著不同的领域，并研究·大脑中反馈的计算使用以及生成性和因果模型如何提高视觉推理算法的稳健性(目标1)·稳健的视觉推理如何受到自然图像获取的动力学的影响(目标2)·稳健的视觉推理如何受到根据神经生物学测量确定的皮质前变换的影响(目标3)。稳健性的概念与泛化的概念密切相关，即处理与以前遇到的情况不同的情况(或任务)的能力。概括的能力还允许在现实和动态的视觉问题上的任务灵活性，这将是新供资期间所有项目的共同主题。所有项目都从儿童权利中心内部密切和富有成效的跨学科互动中受益匪浅，从而产生了激动人心的新假设，这些假设将在第二个资助期进行研究。

The Collaborative Research Centre (CRC) 1233 brings together leading researchers in machine learning, computer vision and systems neuroscience to uncover the computational principles underlying robust vision. Vision plays a key role for animals and humans to achieve a reliable correspondence between the brain's internal model of the world and the external surroundings. It relies on the sensation and interpretation of meaningful patterns implicit in the luminance signals distributed across the entire image. Understanding the principles and algorithms that facilitate robust vision of these patterns plays a fundamental role in understanding biological vision. Big strides have been made in computer vision over recent years with algorithms being tested on increasingly challenging problems, including real-world applications such as autonomous car driving. However, human vision still clearly excels in robustness. The Collaborative Research Centre (CRC) leverages the successes in computer vision and seeks to advance our understanding of the robustness of vision, both in biology and in machines. In close collaborations between computational researchers and neuroscientists, our goal is to uncover the principles of robust vision and to identify its neural basis in the mammalian brain. In particular, the CRC focuses on areas where the neurobiology of vision prominently diverges from current machine vision algorithms and studies• the computational use of feedback in the brain and how generative and causal modelling can improve the robustness of visual inference algorithms (Aim 1)• how robust visual inference is affected by the dynamics of natural image acquisition (Aim 2)• how robust visual inference is affected by pre-cortical transformations as determined from neurobiological measurements (Aim 3).The notion of robustness is tightly linked to the notion of generalisation, i.e. the ability to handle situations (or tasks) that differ from previously encountered situations. The ability to generalize allows also for task flexibility in realistic and dynamic vision problems, which will be a convergent theme of all projects in the new funding period. All projects have strongly benefited from the close and productive interdisciplinary interactions within the CRC leading to exciting new hypotheses that will be investigated in the second funding period.