Fully Dynamic Inference With Deep Neural Networks

• Published in: IEEE transactions on emerging topics in computing Vol. 10; no. 2; pp. 962 - 972
• Main Authors: Xia, Wenhan, Yin, Hongxu, Dai, Xiaoliang, Jha, Niraj K.
• Format: Journal Article
• Language: English
• Published: New York IEEE 01.04.2022; The Institute of Electrical and Electronics Engineers, Inc. (IEEE)
• Subjects: Accuracy; Adaptation models; Artificial neural networks; Autonomous cars; Channels; Computational efficiency; Computational modeling; Computer architecture; Computing costs; Conditional computation; Datasets; deep learning; dynamic execution; dynamic inference; Floating point arithmetic; Inference; model compression; Network latency; Neural networks; Neurons; Quantization (signal); Task analysis
• ISSN: 2168-6750; 2168-6750
• DOI: 10.1109/TETC.2021.3056031

Modern deep neural networks are powerful and widely applicable models that extract task-relevant information through multi-level abstraction. Their cross-domain success, however, is often achieved at the expense of computational cost, high memory bandwidth, and long inference latency, which prevents their deployment in resource-constrained and time-sensitive scenarios, such as edge-side inference and self-driving cars. While recently developed methods for creating efficient deep neural networks are making their real-world deployment more feasible by reducing model size, they do not fully exploit input properties on a per-instance basis to maximize computational efficiency and task accuracy. In particular, most existing methods typically use a one-size-fits-all approach that identically processes all inputs. Motivated by the fact that different images require different feature embeddings to be accurately classified, we propose a fully dynamic paradigm that imparts deep convolutional neural networks with hierarchical inference dynamics at the level of layers and individual convolutional filters/channels. Two compact networks, called Layer-Net (L-Net) and Channel-Net (C-Net), predict on a per-instance basis which layers or filters/channels are redundant and therefore should be skipped. L-Net and C-Net also learn how to scale retained computation outputs to maximize task accuracy. By integrating L-Net and C-Net into a joint design framework, called LC-Net, we consistently outperform state-of-the-art dynamic frameworks with respect to both efficiency and classification accuracy. On the CIFAR-10 dataset, LC-Net results in up to <inline-formula><tex-math notation="LaTeX">11.9\times</tex-math> <mml:math><mml:mrow><mml:mn>11</mml:mn><mml:mo>.</mml:mo><mml:mn>9</mml:mn><mml:mo>×</mml:mo></mml:mrow></mml:math><inline-graphic xlink:href="xia-ieq1-3056031.gif"/> </inline-formula> fewer floating-point operations (FLOPs) and up to 3.3 percent higher accuracy compared to other dynamic inference methods. On the ImageNet dataset, LC-Net achieves up to <inline-formula><tex-math notation="LaTeX">1.4\times</tex-math> <mml:math><mml:mrow><mml:mn>1</mml:mn><mml:mo>.</mml:mo><mml:mn>4</mml:mn><mml:mo>×</mml:mo></mml:mrow></mml:math><inline-graphic xlink:href="xia-ieq2-3056031.gif"/> </inline-formula> fewer FLOPs and up to 4.6 percent higher Top-1 accuracy than the other methods.