Framework for benchmarking models of visual cortex function

神经元之间的通信使用短暂的电压尖峰，导致神经递质的释放。大脑编码关于一个人的环境、记忆和计划的信息，作为涉及许多神经元的协调的尖峰模式。脉冲已经从大脑中记录了50多年，我们知道许多关于它们如何与刺激特性等相关的细节；然而，我们对这些细节如何相互作用产生复杂的行为只有一个粗略的了解。计算模型对于理解这样复杂的系统是必不可少的。模型已经在神经科学中使用了100多年，但从来没有可能开发出能够产生复杂行为的模型。然而，随着深度网络的出现，现在有可能产生越来越复杂的行为；深度网络似乎是产生复杂大脑功能的大规模模型的良好起点。这项研究计划寻求开发这样的模型，从深层网络开始，迭代地纳入生物机制的模型，并测试每种机制是否使网络的内部表示和行为在生物学上更接近现实。研究将特别集中在这一方向上最基本的一步，即开发一套严格的基准测试，将复杂的模型与生物的大脑进行比较。这将用小鼠的大脑来完成，因为关于小鼠大脑的详细信息是公开的，而且因为小鼠大脑的小尺寸将有助于通过不同的模型变化进行快速迭代，以澄清哪些神经机制对神经信息处理做出了最大贡献。具体目标是1)开发一个数据驱动的网络体系结构，允许在深层网络的部分和小鼠大脑之间进行非常具体的比较；2)开发一个现实的虚拟环境，以评估神经机制在小鼠的生活和生存中所扮演的角色；以及3)开发衡量标准，以比较模型和复杂的小鼠大脑活动的公开数据集。这项研究将允许对复杂的大脑模型进行系统和彻底的评估，从而识别最有力地解释尖峰模式和行为的神经回路属性。在小鼠身上的这项工作结果将提供有关人脑功能的重要线索，然后可以在更复杂的模型中进行测试和改进。这项研究将有助于对人类认知的更准确的理解，这可能使新的量化方法能够解决心理学、神经学和教育学中的广泛问题。这项研究还将系统地确定先进人工智能的有用机制。作为这项研究计划的一部分接受培训的HQP将在神经科学和高级人工智能领域的未来职业生涯中处于独特的地位。

Neurons communicate with each other using brief voltage spikes that cause the release of neurotransmitters. The brain encodes information about a person's surroundings, memories, and plans, as coordinated patterns of spikes that involve many neurons. Spikes have been recorded from the brain for over 50 years, and we know many details about how they correlate with stimulus properties, etc.; however, we have only a rough idea of how these details interact to produce complex behaviour. Computational models are essential for understanding such complex systems. Models have been used in neuroscience for over 100 years, but it has never been possible to develop models that produce sophisticated behaviour. However, with the advent of deep networks, it is now possible to produce increasingly sophisticated behaviour; deep networks appear to be good starting points for producing large-scale models of complex brain function. This research program seeks to develop such models, by beginning with deep networks, iteratively incorporating models of biological mechanisms, and testing whether each mechanism makes the networks' internal representations and behaviour more biologically realistic.  The research will focus particularly on the most fundamental step in this direction, which is developing a rigorous suite of benchmark tests to compare sophisticated models to the  brains of living creatures. This will be done with the mouse brain, because detailed information about the mouse brain is publicly available, and because the small size of the mouse brain will facilitate rapid iteration through different model variations, to clarify which neural mechanisms contribute most to neural information processing. The specific objectives are to 1) develop a data-driven network architecture that allows very specific comparisons between parts of deep networks and the mouse brain; 2) develop a realistic virtual environment to evaluate neural mechanisms in terms of their roles in the life and survival of the mouse; and 3) develop metrics to compare models with complex public datasets of mouse brain activity. This research will allow a systematic and thorough assessment of sophisticated brain models, leading to identification of the neural circuit properties that most strongly account for spike patterns and behaviour. The results of this work in the mouse will provide important hints about the function of the human brain, which can then be tested and refined in more complex models. The research will contribute to a more precise understanding of human cognition, which may enable new quantitative approaches to a wide range of problems in psychology, neurology, and education. The research will also systematically identify useful mechanisms for advanced artificial intelligence. HQP trained as part of this research program will be uniquely positioned for future careers in neuroscience and advanced artificial intelligence.