EVENT DOSSIER
The researchers proposed the Deep Microcompression (DMC) scheme, aimed at providing deep learning inference support for microcontrollers with limited resources. This scheme integrates structured pruning, quantization-aware training, and fixed-length bit packing techniques, achieving a weight compression ratio of 55.8 times on the LeNet-5 model while maintaining an accuracy of 98.77%. The generated code has no external dependency libraries and its latency is determined. On the RP2040 (Cortex-M0+) platform, the binary size of DMC is 3 times smaller than that of TensorFlow Lite while maintaining the same precision. A key breakthrough is that DMC successfully deployed standard convolutional neural networks on the ATmega328P device with only 2KB of SRAM, solving the technical challenge that previously made it impossible to perform CNN inference on this device.
本文提出了一种面向裸机微控制器的深度学习推理硬件感知流水线 Deep Microcompression（DMC）。该方案整合了结构化剪枝、量化感知训练及固定长度位打包技术，在 LeNet-5 模型上实现了 55.8 倍的权重压缩比并保持 98.77% 的准确率，生成了无依赖库且延迟确定的代码。在 RP2040（Cortex-M0+）平台上，DMC 的二进制体积较 TensorFlow Lite 减少 3 倍并匹配其精度。关键突破在于，DMC 首次实现了标准卷积神经网络在 ATmega328P 上的部署，该设备受限于仅 2KB SRAM，此前被认为无法进行 CNN 推理。