Brain-inspired deep neural networks such as Convolutional Neural Network (CNN) have shown great potential in solving difficult cognitive problems such as object recognition and classification. However, such architectures have high computational energy demand and sensitivity to variation effects, making them inapplicable for energy-constrained embedded learning platforms. To address this issue, we propose a learning and mapping approach that utilizes approximate computing during early design phases for a layer-wise pruning and fault tolerant weight mapping scheme of reliable and energy-efficient CNNs. In the proposed approach, approximate CNN is prepared first by layer-wise pruning of approximable neurons, which have high error tolerance margins using a two-level approximate learning methodology. Then, the pruned network is retrained to improve its accuracy by fine-tuning the weight values. Finally, a fault-tolerant layer-wise neural weight mapping scheme is adopted to aggressively reduce memory operating voltage when loading the weights of error resilient layers for energy-efficiency. Thus, the combination of approximate learning and fault tolerance aware memory operating voltage downscaling techniques enable us to implement robust and energy-efficient approximate inference engine for CNN applications. Simulation results show that the proposed fault tolerant and approximate learning approach can improve the energy-efficiency of CNN inference engines by more than 50% with less than 5% reduction in classification accuracy. Additionally, more than 26% energy-saving is achieved by using the proposed layer-wise mapping-based cache memory operating voltage down-scaling.

中文翻译：

---

节能神经形态系统的近似学习和容错映射

卷积神经网络 (CNN) 等受大脑启发的深度神经网络在解决诸如物体识别和分类等困难的认知问题方面显示出巨大的潜力。然而，这种架构具有很高的计算能量需求和对变化效应的敏感性，使其不适用于能量受限的嵌入式学习平台。为了解决这个问题，我们提出了一种学习和映射方法，该方法利用近似计算在早期设计阶段，用于可靠和节能的 CNN 的逐层修剪和容错权重映射方案。在所提出的方法中，近似 CNN 首先是通过对近似神经元进行逐层修剪来准备的，这些神经元使用两级近似学习方法具有较高的容错余量。然后，对修剪后的网络进行重新训练，以通过微调权重值来提高其准确性。最后，在加载错误弹性层的权重以提高能效时，采用容错的逐层神经权重映射方案来积极降低内存工作电压。因此，近似学习和容错感知内存操作电压缩减技术的结合使我们能够为 CNN 应用程序实现稳健且节能的近似推理引擎。仿真结果表明，所提出的容错和近似学习方法可以将 CNN 推理引擎的能效提高 50% 以上，而分类精度降低不到 5%。此外，通过使用所提出的基于分层映射的高速缓存存储器工作电压缩减，可实现超过 26% 的节能。