Materials and devices as solutions to computational problems in machine learning
Materials and devices as solutions to computational problems in machine learning
复制标题
DOI：
10.1038/s41928-023-00977-1
复制
发表时间：
2023-07
影响因子：
34.3
通讯作者：
N. Tye;Stephan Hofmann;Phillip Stanley-Marbell
中科院分区：
文献类型：
--
作者：
N. Tye;Stephan Hofmann;Phillip Stanley-Marbell
机器学习的发展，加上传统数字计算的接近极限，正在推动寻找替代和补充计算形式，但主流计算系统很少采用新颖的设备。这种计算机技术的发展需要计算设备和计算机体系结构的进步。然而，设备社区和计算机体系结构社区之间存在脱节，这限制了进展。在这里，我们重点探讨机器学习硬件加速器上的这种脱节。我们认为，计算问题与材料和设备属性的直接映射提供了一条强有力的前进道路。我们研究了已成功应用于计算问题解决方案的新型材料和设备：用于矩阵向量乘法的非易失性存储器、用于随机计算的磁隧道结和用于可重构逻辑的电阻存储器。我们还提出了一些指标，以促进机器学习任务的不同解决方案之间的比较，并突出显示可能使用新颖材料和设备的应用。
The growth of machine learning, combined with the approaching limits of conventional digital computing, are driving a search for alternative and complementary forms of computation, but few novel devices have been adopted by mainstream computing systems. The development of such computer technology requires advances in both computational devices and computer architectures. However, a disconnect exists between the device community and the computer architecture community, which limits progress. Here we explore this disconnect with a focus on machine learning hardware accelerators. We argue that the direct mapping of computational problems to materials and device properties provides a powerful route forwards. We examine novel materials and devices that have been successfully applied as solutions to computational problems: non-volatile memories for matrix-vector multiplication, magnetic tunnel junctions for stochastic computing and resistive memory for reconfigurable logic. We also propose metrics to facilitate comparisons between different solutions to machine learning tasks and highlight applications where novel materials and devices could potentially be of use.