COG

COG数据集是由纽约大学神经科学中心和Google Brain共同开发的，旨在模拟人类和动物的视觉推理和记忆实验。该数据集通过程序语言生成一系列任务三元组：图像序列、语言指令和正确答案序列，涵盖广泛的视觉推理任务，并要求对文本的语义理解、图像的视觉感知以及工作记忆的运用。数据集设计灵活，可根据不同难度进行调整，适用于测试现代深度学习架构在视觉和逻辑推理及记忆方面的能力。COG数据集主要应用于研究时间变化的视觉刺激中的推理问题，为开发具有强大和灵活推理能力的系统提供基准。

The COG dataset was co-developed by the Center for Neuroscience at New York University and Google Brain, aiming to simulate visual reasoning and memory experiments on humans and animals. This dataset generates a series of task triplets via programming: image sequences, language instructions, and correct answer sequences. It covers a broad spectrum of visual reasoning tasks, and requires the application of textual semantic understanding, visual perception of images, and working memory. The dataset features a flexible design that can be adjusted for varying difficulty levels, making it suitable for testing the capabilities of modern deep learning architectures in visual and logical reasoning as well as memory. The COG dataset is primarily applied to research on reasoning problems in time-varying visual stimuli, serving as a benchmark for developing systems with robust and flexible reasoning capabilities.