In energy-harvesting intermittent computing systems, balancing power constraints with the need for timely and accurate inference remains a critical challenge. Existing methods often sacrifice significant accuracy or fail to adapt effectively to fluctuating power conditions. This paper presents DualAdaptNet, a power- and deadline-aware neural network architecture that dynamically adapts both its width and depth to ensure reliable inference under variable power conditions. Additionally, a runtime scheduling method is introduced to select an appropriate sub-network configuration based on real-time energy-harvesting conditions and system deadlines. Experimental results on the MNIST dataset demonstrate that our approach completes up to 7.0% more inference tasks within a specified deadline, while also improving average accuracy by 15.4% compared to the state-of-the-art.

Distributed inference is a popular approach for efficient DNN inference at the edge. However, traditional Static and Dynamic DNNs are not distribution-friendly, causing system reliability and adaptability issues. In this paper, we introduce Fluid Dynamic DNNs (Fluid DyDNNs), tailored for distributed inference. Distinct from Static and Dynamic DNNs, Fluid DyDNNs utilize a novel nested incremental training algorithm to enable independent and combined operation of its sub-networks, enhancing system reliability and adaptability. Evaluation on embedded Arm CPUs with a DNN model and the MNIST dataset, shows that in scenarios of single device failure, Fluid Dy DNNs ensure continued inference, whereas Static and Dynamic DNNs fail. When devices are fully operational, Fluid DyDNNs can operate in either a High-Accuracy mode and achieve comparable accuracy with Static DNNs, or in a High-Throughput mode and achieve 2.5x and 2x throughput compared with Static and Dynamic DNNs, respectively.

To facilitate large-scale deployment of convolutional networks, integer-arithmetic-only inference has been demonstrated effective, which not only reduces computational cost but also ensures cross-platform consistency. However, previous studies on integer networks usually report a decline in the inference accuracy, given the same number of parameters as floating-point-number (FPN) networks. In this paper, we propose to finetune and quantize a well-trained FPN convolutional network to obtain an integer convolutional network. Our key idea is to adjust the upper bound of a bounded rectified linear unit (ReLU), which replaces the normal ReLU and effectively controls the dynamic range of activations. Based on the tradeoff between learning ability and quantization error of networks, we managed to preserve full accuracy after quantization and obtain efficient integer networks. Our experiments on ResNet for image classification demonstrate that our 8-bit integer networks achieve state-of-the-art performance compared with Google's TensorFlow and NVIDIA's TensorRT. Moreover, we experiment on VDSR for image super-resolution and on VRCNN for compression artifact reduction, both of which serve regression tasks that natively require high inference accuracy. Besides ensuring the equivalent performance as the corresponding FPN networks, our integer networks have only 1/4 memory cost and run 2× faster on GPUs.