Large-scale neural models of cognitive function

我的实验室开发了一种方法来模拟认知行为，使用生物学上合理的刺突神经元，从而产生了目前世界上最大的功能性大脑模型——Spaun。最新版本的Spaun拥有超过660万个神经元和200亿个连接，能够执行各种各样的任务。这个模型展示了我的实验室开发的两个理论，神经工程框架（NEF）和语义指针架构（SPA）；总之，这些概述了在尖峰神经网络中构建认知模型的一般方法和架构，并实现了一种独特的生物认知方法。我们现在遇到的最重要的挑战，也是本提案的重点，是扩大理论和模拟的规模，以在更苛刻的环境和更多的生物细节中解决认知任务。我们已经确定了需要解决的三个主要挑战，以保持我们在该领域的领导地位。首先，在捕捉心理功能的同时，还没有一个小组在大规模认知模型中使用高度逼真的单神经元模型（例如，隔间模型、基于电导的模型）。这样做为更好地描述生物和心理特性之间的关系提供了希望，使我们能够测试新的医疗干预措施（例如药物，深部脑刺激等）。第二，生物系统非常擅长处理时间和空间上的连续性。目前的理论没有充分处理神经网络如何表示和处理这种连续性。我们将继续开发和利用空间语义指针（ssp）和勒让德记忆单元（LMU），这是我们为此目的而开发的。第三，理解和利用适应在像大脑这样一个庞大、复杂的功能系统中所起的关键作用是困难的。挑战来自于确保系统在适应过程中保持稳定，并在认知层面上推广跨任务的学习行为。解决这三个广泛的挑战将使我们能够构建下一代的“全脑”模型。这项研究将进一步推进这些理论框架，在生物学上合理的模型中整合感知、运动和认知行为。本研究将显著提高NEF和SPA模型的适应性和功能，并将模型与医学干预相结合，有助于神经系统损伤的诊断、预防和修复，具有重要的实际意义。此外，了解自然界为现实世界难题找到的独特解决方案，将为智能系统、模式识别和决策应用中的创新工程解决方案铺平道路。

My lab has developed methods for simulating cognitive behaviour using biologically plausible spiking neurons that have resulted in what is currently the world's largest functional brain model, Spaun. The most recent version of Spaun has over 6.6 million neurons with 20 billion connections, and is able to perform a wide variety of tasks. This model demonstrates two theories developed in my lab, the Neural Engineering Framework (NEF) and the Semantic Pointer Architecture (SPA); together, these outline a general method and architecture for constructing cognitive models in spiking neural networks, and realize a unique approach to biological cognition. The overriding challenge that we now encounter, and the focus of this proposal, is scaling up both the theory and the simulations to tackle cognitive tasks in more demanding circumstances and with greater biological detail. We have identified three main challenges that we need to address to retain our position as leaders in the field. First, no group has yet employed highly realistic single neuron models (e.g. compartmental, conductance based models) in large-scale cognitive models while capturing psychological functions. Doing so provides the promise of better characterizing the relationship between biological and psychological properties allowing us to test new medical interventions (e.g. drugs, deep brain stimulation, etc.). Second, biological systems are extremely adept at dealing with continuity in time and in space. Current theory does not adequately deal with how neural networks represent and process such continuity. We will continue to develop and exploit Spatial Semantic Pointers (SSPs) and the Legendre Memory Unit (LMU), which we developed for such purposes. Third, understanding and exploiting the critical role that adaptation plays in a large, complex, functioning system like the brain, is difficult. Challenges stem from ensuring that the system remains stable during adaptation and generalizing learned behaviours across tasks at the cognitive level. Addressing these three broad challenges will allow us to construct the next generation of `whole-brain' models. This research will further advance these theoretical frameworks for integrating perceptual, motor, and cognitive behaviour in biologically plausible models. The proposed research will significantly improve the adaptability and functionality of NEF and SPA models, and have the important practical consequence of connecting models to medical interventions that can help with the diagnosis, prevention, and repair of damage to neural systems. In addition, understanding the unique solutions that nature has found for difficult real-world problems will pave the way for innovative engineering solutions in intelligent systems, pattern recognition, and decision-making applications.