Spatial-Temporal Hybrid Neural Network With Computing-in-Memory Architecture
Spatial-Temporal Hybrid Neural Network With Computing-in-Memory Architecture
复制标题
DOI：
10.1109/tcsi.2021.3071956
复制
发表时间：
2021-04
期刊：
影响因子：
--
通讯作者：
Kangjun Bai;Lingjia Liu;Y. Yi
中科院分区：
文献类型：
--
作者：
Kangjun Bai;Lingjia Liu;Y. Yi
深度学习在许多实际应用中取得了前所未有的成功。然而，由于复杂的基于梯度的学习算法的需要和突触权重存储所需的高存储带宽，特别是在当今数据密集型的环境中，动态学习算法给高效的硬件实现带来了困难。内存计算(CIM)策略已经成为在硅中实现高能效的神经形态应用的替代方案，减少了神经计算所需的资源和能量。在这项工作中，我们开发了一种基于CIM的时空混合神经网络(STHNN)，它具有独特的学习算法。具体地说，我们集成了多层感知器和基于递归的延迟动力系统，使得网络在处理空间和时间域信息的同时成为线性可分的，更好的是，通过CIM结构减少了存储带宽和硬件开销。该样机采用180 nm CMOS工艺，采用全模拟器件制作，对手写字母字符的片上平均分类准确率高达86.9%，功耗为33 mW。此外，通过手写数字数据库和射频指纹数据集，基于软件的数值评估分别提供了1.6美元到9.8倍和1.9美元到4.4倍的加速比，而与尖端的DL方法相比，其分类精度没有显著下降。
Deep learning (DL) has gained unprecedented success in many real-world applications. However, DL poses difficulties for efficient hardware implementation due to the needs of a complex gradient-based learning algorithm and the required high memory bandwidth for synaptic weight storage, especially in today’s data-intensive environment. Computing-in-memory (CIM) strategies have emerged as an alternative for realizing energy-efficient neuromorphic applications in silicon, reducing resources and energy required for neural computations. In this work, we exploit a CIM-based spatial-temporal hybrid neural network (STHNN) with a unique learning algorithm. To be specific, we integrate both multilayer perceptron and recurrent-based delay-dynamical system, making the network becomes linear separable while processing information in both spatial and temporal domains, better yet, reducing the memory bandwidth and hardware overhead through the CIM architecture. The prototype fabricated in 180 nm CMOS process is built of fully-analog components, yielding an average on-chip classification accuracy up to 86.9% on handprinted alphabet characters with a power consumption of 33 mW. Beyond that, through the handwritten digit database and the radio frequency fingerprinting dataset, software-based numerical evaluations offer $1.6 -to- 9.8 \times $ and $1.9 {-to-} 4.4 \times $ speedup, respectively, without significantly degrading its classification accuracy compared to the cutting-edge DL approaches.