Memory Consolidation in Memristive Systems

本研究将计算神经科学理论与新的记忆技术相结合，以开发具有长寿命记忆巩固和回忆功能的脑启发神经系统。最初的兴趣来自于识别当前机器学习理论的局限性，即冯-诺伊曼体系结构的低效率和它们采用的静态存储方法。该项目的主要目标是实现一种新技术，该技术具有始终在线、始终学习的能力，并且可以集成到其他物理和可编程系统中，与物联网社会一起流动。工程方法应采用自下而上的方法来开发硅内神经电路。我的目标是建立在“记忆融合”的概念上，实现具有元可塑性的人工突触模型，从而能够动态地存储具有内在时间参考的信息。这种技术的理论建模将在Python中完成，制造的设备将使用Arc Instruments分析工具进行测试。成功的模型应在数量上进行整合，以开发具有巩固和回忆功能的主动记忆装置。量化个体记忆的理论工作是实现存储中最大模块化层次的重要任务。由部分相同信息组成的多重记忆应该与相应的预先存在的突触相关联，其效力应该得到加强。这种设备中记忆巩固的拓扑结构最令人感兴趣，因为它可以在信息输入和推理之间建立新的桥梁。贝叶斯统计等数学模型将用于研究上述硬件内存召回方法。系统性能可以通过随机组合突触信息来创建和存储新的记忆，从而设计出新的“创造性”方法。这种改进可以是人工自动化系统解决多概念问题能力的重大飞跃。这项研究将与智能社会和更普遍的人工智能的趋势完全同步。潜在的影响是巨大的，因为成功的结果可以为计算开辟新的途径，最终克服当前内存操作的静态限制。

Key objectivesThis research will combine computational neuroscience theory with novel memristive technologies for the development of brain-inspired neural systems that exhibit long-life memory consolidation and recall. Initial interest has risen from identifying the limitations of current machine-learning theory with respect to both inefficiencies of von-Neumann architecture and the static memory methods they employ. The main goal for this project is the realisation of a new technology that exhibits always-on, always-learning capabilities and can be integrated to other physical and programmable systems, in flow with the Internet of Things society.Engineering MethodologyA bottom-up approach shall be examined to develop neural circuits in-Silico. I aim to build on the "memristive fuse" concept to realise artificial synaptic models that exhibit meta-plasticity and are therefore able to dynamically store information with an intrinsic time reference. Theoretical modelling of such technology will be done in Python and fabricated devices will be tested using Arc Instruments analysis tools.Successful models shall be integrated in numbers, to develop active memory devices exhibiting consolidation and recall. Theoretical work on quantifying individual memories is an essential task to achieve maximum modular hierarchy in storage. Multiple memories consisting partially of the same information should be associated to corresponding pre-existing synapses, whose efficacies should be re-enforced. The topology of memory consolidation in such devices is of most interest since it can create new bridges between information input and reasoning.Mathematical models such as Bayesian statistics will be used to study the mentioned hardware memory recall methods. System performance can be designed to exhibit new "creative" methods by stochastically combining synaptic information to create and store new memories. Such improvement can be a significantly leap in the multi-concept problem-solving abilities of artificial automated systems.ApplicationsThis research will be in full sync with the trend towards a smart society and more general artificial intelligence. The potential impact is vast since successful results can open new paths for computation, finally overcoming the static limitations of current memory manipulation.