Analyzing inference robustness of RRAM synaptic array in low-precision neural network

In this work, we investigate the robustness of 1-transistor-1-resistor (1T1R) synaptic array to implement a low-precision neural network. The experimental results on 1 kb HfOx-based RRAM array show a large on/off ratio (i.e. > 105×) and 5 stable resistance states can be reliably achieved with 10× window between adjacent two states. As the RRAM has the resistance drift over time under read voltage stress, the impact of read disturbance occurred in 1T1R synaptic array on the neural network classification accuracy is analyzed with the RRAM compact model fitted with experimental data. The simulation results of a single-layer perceptron with compressed MNIST dataset indicate that 1) more stable multi-level states are desired to have higher mapping capability of weights, thus achieving a higher initial classification accuracy; 2) good mapping strategies that avoid the read disturbance-induced sign change on the most significant weight levels are very important to mitigate the classification accuracy loss.

[1]  Yoshua Bengio,et al.  BinaryConnect: Training Deep Neural Networks with binary weights during propagations , 2015, NIPS.

[2]  Shimeng Yu,et al.  Scaling-up resistive synaptic arrays for neuro-inspired architecture: Challenges and prospect , 2015, 2015 IEEE International Electron Devices Meeting (IEDM).

[3]  Igor Carron,et al.  XNOR-Net: ImageNet Classification Using Binary Convolutional Neural Networks , 2016 .

[4]  Guigang Zhang,et al.  Deep Learning , 2016, Int. J. Semantic Comput..

[5]  Geoffrey E. Hinton,et al.  ImageNet classification with deep convolutional neural networks , 2012, Commun. ACM.

[6]  Frederick T. Chen,et al.  Low power and high speed bipolar switching with a thin reactive Ti buffer layer in robust HfO2 based RRAM , 2008, 2008 IEEE International Electron Devices Meeting.

[7]  R. Jordan,et al.  NVM neuromorphic core with 64k-cell (256-by-256) phase change memory synaptic array with on-chip neuron circuits for continuous in-situ learning , 2015, 2015 IEEE International Electron Devices Meeting (IEDM).

[8]  Shimeng Yu,et al.  Compact Modeling of RRAM Devices and Its Applications in 1T1R and 1S1R Array Design , 2015, IEEE Transactions on Electron Devices.

[9]  Zhiwei Li,et al.  Binary neural network with 16 Mb RRAM macro chip for classification and online training , 2016, 2016 IEEE International Electron Devices Meeting (IEDM).

[10]  E. Vianello,et al.  HfO2-Based OxRAM Devices as Synapses for Convolutional Neural Networks , 2015, IEEE Transactions on Electron Devices.