Complex Data Processing with Novel Memristor Array

Abstract

최근 deep learning의 대두로 다양한 data들이 축적되고 학습에 사용되었다. Big data는 그 구조가 더욱 다양화되고 복잡해지면서 기존 하드웨어로 처리하기 힘든 complex data 가 등장했다. Complex data의 예시로는 Sequential data, graph data 가 있다. Sequential data는 현재 스테이트가 인풋 히스토리를 반영하면서 그 패턴이 일정하지 않고 예측하기 힘든 특성이 있다. 그래프 타입의 데이터는 주체와 주체간의 연결성들을 다루기에 vector 형태로 표현되기 어려워, 기존 하드웨어 구조에서 처리하기 힘들다는 문제가 있다. 이런 복잡한 data를 처리하기 위해서는 novel data processing technique이 요구된다. 본 연구의 첫번째 파트에서는, 효과적인 시퀀셜 데이터 처리를 위해서 멤리스터, 리지스터, 캐패시터를 이용해 temporal kernel 을 구성하였다. 전체적인 컴퓨팅 스킴은 conventional reservoir system 과 동일하여 input 이 temporal kernel 에서 처리된 데이터가 멤리스터에 저장된다. 이후 이러한 멤리스터 컨덕턴스 벡터를 인풋으로 readout network 를 학습시킨다. 유닛셀은 멤리스터가 리지스터, 캐패시터와 직렬연결되어 있고 리지스터와 캐패시터는 서로 병렬 연결되어 있는 1M1R1C 구조를 가진다. 1M1R1C kernel은 R, C 조절을 통해 다양한 time constant 를 가질 수 있어 다양한 상황에 적용가능하다는 장점이 있다. 본 연구에서는 1M1R1C 기반 MNIST recognition에서 높은 에너지 효율과 빠른 처리속도로 높은 정확도 (90 %)를 보였다. 한편 1M1R1C kernel은 시간 상수가 매우 다른 ultrasound, electrocardiogram 기반 medical diagnosis에도 적용되어 1 ~ 10 MHz 의 넓은 주파수 영역에서 성공적으로 task를 수행하였다. 본 연구의 두번째 파트에서는, 자가정류 멤리스터 어레이를 이용해 비유클리드 그래프를 처리하는 방법이 다뤄진다. 비유클리드 그래프에서는 유사도를 구할 수 없어 그래프 임베딩 등의 복잡한 전처리 과정이 요구되며 그 과정에서 정보의 손실도 발생한다. 본 연구에서는 비유클리드 그래프를 벡터화하지 않고 그 본래 데이터 그대로 맵핑하고 분석하는 방법을 제안한다.

Recently, with the remarkable development of deep learning, various data have been accumulated. As the structure of big data becomes more diversified and complex, complex data that is difficult to process with existing hardware has emerged. Examples of complex data include sequential data and graph data. Sequential data has characteristics that the current state reflects the input history and the pattern is not constant and difficult to predict. Graph-type data is difficult to be expressed in vector form since graphical data includes the connections between entities. To process such complex data, novel data processing techniques are required. In the first part of this study, a method for processing time-series data with a nonvolatile memristor is proposed. Recent advances in physical reservoir computing, which is a type of temporal kernel, have made it possible to perform complicated timing-related tasks using a linear classifier. However, the fixed reservoir dynamics in previous studies have limited application fields. In this study, temporal kernel computing was implemented with a physical kernel that consisted of a W/HfO2/TiN memristor, a capacitor, and a resistor, in which the kernel dynamics could be arbitrarily controlled by changing the circuit parameters. After the capability of the temporal kernel to identify the static MNIST data was proven, the system was adopted to recognize the sequential data, ultrasound (malignancy of lesions), and electrocardiogram (arrhythmia), that had a significantly different time constant (10 7 vs. 1 s). The suggested system feasibly performed the tasks by simply varying the capacitance and resistance. These functionalities demonstrate the high adaptability of the present temporal kernel compared to the previous ones. In the second part of this study, a method for processing non-Euclidean graphs using self-rectifying memristor arrays is proposed. Many big data have interconnected and dynamic graph structures growing over time. Analyzing these graphical data requires identifying the hidden relationship between the nodes in the graphs, which has conventionally been achieved by finding the effective similarity. However, graphs are generally non-Euclidean, which does not allow finding it. In this study, the non-Euclidean graphs were mapped to a specific crossbar array (CBA) composed of the self-rectifying memristors and metal cells at the diagonal positions. When all bit lines of CBA are connected to the ground, the sneak current is suppressed, and CBA can be used to search for adjacent nodes. When a single bit line is connected to the ground, the sneak current, an intrinsic physical property of the CBA, allows for identifying the similarity function. Sneak current-based similarity function indicates the distance between nodes, the probability that unconnected nodes will be connected in the future, connectivity between communities, and cortical connections in a brain. This work demonstrates the physical calculation methods applied to various graphical problems using the CBA composed of the self-rectifying-memristor based on the HfO2 switching layer. Moreover, such applications suffer less from the memristors' inherent issues related to their stochastic nature.