Bit Error Robustness for Energy-Efficient DNN Accelerators
arXiv:2006.13977
Abstract
Deep neural network (DNN) accelerators received considerable attention in past years due to saved energy compared to mainstream hardware. Low-voltage operation of DNN accelerators allows to further reduce energy consumption significantly, however, causes bit-level failures in the memory storing the quantized DNN weights. In this paper, we show that a combination of robust fixed-point quantization, weight clipping, and random bit error training (RandBET) improves robustness against random bit errors in (quantized) DNN weights significantly. This leads to high energy savings from both low-voltage operation as well as low-precision quantization. Our approach generalizes across operating voltages and accelerators, as demonstrated on bit errors from profiled SRAM arrays. We also discuss why weight clipping alone is already a quite effective way to achieve robustness against bit errors. Moreover, we specifically discuss the involved trade-offs regarding accuracy, robustness and precision: Without losing more than 1% in accuracy compared to a normally trained 8-bit DNN, we can reduce energy consumption on CIFAR-10 by 20%. Higher energy savings of, e.g., 30%, are possible at the cost of 2.5% accuracy, even for 4-bit DNNs.
References in corpus (36)
- Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift
- Fashion-MNIST: a Novel Image Dataset for Benchmarking Machine Learning Algorithms
- Intriguing properties of neural networks
- Improved Regularization of Convolutional Neural Networks with Cutout
- ZOO: Zeroth Order Optimization based Black-box Attacks to Deep Neural Networks without Training Substitute Models
- DoReFa-Net: Training Low Bitwidth Convolutional Neural Networks with Low Bitwidth Gradients
- Quantized Neural Networks: Training Neural Networks with Low Precision Weights and Activations
- Delving into Transferable Adversarial Examples and Black-box Attacks
- Quantizing deep convolutional networks for efficient inference: A whitepaper
- AutoAugment: Learning Augmentation Policies from Data
- PACT: Parameterized Clipping Activation for Quantized Neural Networks
- Provable defenses against adversarial examples via the convex outer adversarial polytope
- Fixed Point Quantization of Deep Convolutional Networks
- Reliable evaluation of adversarial robustness with an ensemble of diverse parameter-free attacks
- Robustness May Be at Odds with Accuracy
- Distributional Smoothing with Virtual Adversarial Training
- Efficient Neural Network Robustness Certification with General Activation Functions
- Benchmarking Neural Network Robustness to Common Corruptions and Perturbations
- On the Effectiveness of Interval Bound Propagation for Training Verifiably Robust Models
- Learning with a Strong Adversary
- A Survey on Methods and Theories of Quantized Neural Networks
- Are Labels Required for Improving Adversarial Robustness?
- Resiliency of Deep Neural Networks under Quantization
- Fixup Initialization: Residual Learning Without Normalization
- Training Quantized Nets: A Deeper Understanding
- Improving Robustness Without Sacrificing Accuracy with Patch Gaussian Augmentation
- Deep neural networks are robust to weight binarization and other non-linear distortions
- Adversarial Robustness Against the Union of Multiple Perturbation Models
- Lets keep it simple, Using simple architectures to outperform deeper and more complex architectures
- Batch Normalization is a Cause of Adversarial Vulnerability
- Efficient Processing of Deep Neural Networks: A Tutorial and Survey
- LQ-Nets: Learned Quantization for Highly Accurate and Compact Deep Neural Networks
- Bit Fusion: Bit-Level Dynamically Composable Architecture for Accelerating Deep Neural Networks
- Combinatorial Attacks on Binarized Neural Networks
- Gradient Regularization for Quantization Robustness
- Significance Driven Hybrid 8T-6T SRAM for Energy-Efficient Synaptic Storage in Artificial Neural Networks