Coordination Funds

过去的十年表明，如果联合学习整个视觉数据的处理，绝大多数视觉计算问题都会有更高质量的解决方案：深度学习时代已经在很大程度上取代了以前的顺序方法，例如首先为特定任务设计重要特征，然后学习使用这些特征来分析或分类数据。然而，这种所谓的端到端学习范式将传感器系统记录的图像数据视为学习管道的开始。它忽略了这样一个事实，即图像数据本身是在开发和动态适应传感器系统时具有许多设计选择的上游过程的结果，因此仍然是一种顺序方法，其中传感器系统的设计与数据处理流水线分开。“学会感知”(L2S)研究单元的目标是与神经网络一起对传感器系统的设计参数进行联合优化，以分析结果数据，即开发真正的端到端机器学习方法，产生针对特定应用任务进行优化的系统。因此，L2S项目将对两者进行联合基础研究，使传感器系统自适应以提供有希望的自由度，并对机器学习方法进行联合优化，从而实现传感器系统和网络参数的联合优化。从长远来看，L2S范式将为自适应传感器系统与具有最佳特定任务特性的神经网络的集成开发提供一种新的方法，从而在传感器系统设计中以最小的人工推理实现更高效、更精确的场景分析。

The past decade has shown that a vast majority of visual computing problems admit significantly higher quality solutions if the entire processing of the visual data is learned jointly: The era of Deep Learning has largely replaced previous sequential approaches such as first designing important features for a specific task and subsequently learning to analyze or classify the data using such features. Yet, this so-called end-to-end learning paradigm considers the image data a sensor system records as the beginning of the learning pipeline. It neglects the fact that the image data itself is the result of an upstream process with many design choices in developing and dynamically adapting the sensor system, and thus still remains a sequential approach in which the sensor system is designed separate from the data processing pipeline. The goal of the "Learning to Sense" (L2S) research unit is a joint optimization for design parameters of the sensor system along with the neural network to analyze the resulting data, i.e., developing a true end-to-end machine learning methodology yielding systems that are optimized for an application specific task. Consequently, the L2S project will conduct joint fundamental research on both, making sensor systems adaptive to provide promising degrees of freedom, and on a machine learning methodology that allows for the joint optimization of the resulting sensor system and network parameters. In the long run, the L2S paradigm will provide a new methodology for the integral development of adaptive sensor systems alongside neural networks with optimal task-specific characteristics, which results in a substantially more efficient and more precise scene analysis with minimal manual inference in the sensor system design.