IoT Edge intelligence requires Convolutional Neural Network (CNN) inference to take place in the edge devices itself. ARM big.LITTLE architecture is at the heart of prevalent commercial edge devices. It comprises of single-ISA heterogeneous cores grouped into multiple homogeneous clusters that enable power and performance trade-offs. All cores are expected to be simultaneously employed in inference to attain maximal throughput. However, high communication overhead involved in parallelization of computations from convolution kernels across clusters is detrimental to throughput. We present an alternative framework called Pipe-it that employs pipelined design to split convolutional layers across clusters while limiting parallelization of their respective kernels to the assigned cluster. We develop a performance-prediction model that utilizes only the convolutional layer descriptors to predict the execution time of each layer individually on all permitted core configurations (type and count). Pipe-it then exploits the predictions to create a balanced pipeline using an efficient design space exploration algorithm. Pipe-it on average results in a 39% higher throughput than the highest antecedent throughput.

Current sensor technologies facilitate device-free and non-invasive monitoring of target activities and infrastructures to ensure a safe and inhabitable environment. Device-free techniques for sensing the surrounding environment are an emerging area of research where a target does not need to carry or attach any device to provide information about its motion or the surrounding environment. Consequently, there has been an increasing interest in device-free sensing. Seismic sensors are extremely effective tools for gathering target motion information. In this paper, we provide a comprehensive overview of the seismic sensor-based device-free sensing process and highlight the key techniques within the research field. We classify the existing literature into three categories, viz., (i) target detection, (ii) target localization, and (iii) target identification, and activity recognition. The techniques in each category are divided into multiple subcategories in a structured manner to comprehensively discuss the details. We also discuss the challenges associated with contemporary cutting-edge research and suggest potential solutions.

Human target localization has a plethora of ubiquitous applications, including healthcare and security. Reliable location estimation is challenging in outdoor settings because most targets do not carry, wear, or attach a sensor to their bodies to provide motion information. Seismic and audio sensors can be used extensively to monitor targets in an unobtrusive manner. This paper proposes regression-based methods to localize a human target in an outdoor environment using seismic and audio signatures. The approaches fall into two groups. The first group of approaches applies a late fusion on the audio and seismic modality for target localization. First, a regression algorithm on the seismic signals is applied to estimate the distance of the target from each sensor; then, the target distance information is combined with the audio direction to infer the target location. In contrast to the first group of approaches, the second group of approaches applies an early fusion of audio and seismic modality to directly predict the target location. We compare the proposed methods with multiple state-of-the-arts on different evaluation measures. On 5-fold cross-validation, we achieve a root-mean-localization error of 0.735 and 0.907 meters in an area of 324 meter ² , for dataset-1 and dataset-2, respectively. The best results with proposed approaches show an improvement of 4.68 and 4.57 meters over the best performing state-of-the-art approach for dataset-1 and dataset-2, respectively. We also analyzed the generalization capabilities of the proposed methods. Further, a detailed ablation study on the number of sensors required for localization, feature selection, training data requirement, etc., has been carried out.

Batteryless technology evolved to replace battery technology. Nonvolatile memory (NVM)-based processors were explored to store the program state during a power failure. The energy stored in a capacitor is used for a backup during a power failure. Since the size of a capacitor is fixed and limited, the available energy in a capacitor is also limited and fixed. Thus, the capacitor energy is insufficient to store the entire program state during frequent power failures. This article proposes an architecture that assures safe backup of volatile contents during a power failure under energy constraints. Using a proposed dirty block table (DBT) and writeback queue (WBQ), this work limits the number of dirty blocks in the L1 cache at any given time. We further conducted a set of experiments by varying the parameter sizes to help the user make appropriate design decisions concerning their energy requirements. The proposed architecture decreases energy consumption by 17.56%, the number of writes to NVM by 18.97% at last level cache (LLC), and 10.66% at a main-memory level compared to baseline architecture.

Agriculture is facing bigger challenges in the 21st century due to the scarcity of resources. Artificial intelligence is being integrated with agriculture to cater to people's needs, unlocking fresh avenues for sustainability and innovation. One of the crucial agricultural practices is plant growth monitoring to detect plant stress at an early stage. In the past, there have been preliminary attempts at plant growth monitoring using red–green–blue (RGB) and depth images. The major challenge of this approach is the unavailability of the depth camera at the farmers' end. In this work, we have developed a transformer-based plant age RGB-depth fusion knowledge distillation network (PA-RDFKNet), a multi-to-single modal teacher–student network, that exploits the combined knowledge of RGB-depth pairs at the training time to infer the growth using RGB images alone during test time. The model uses a distillation loss that combines response-based, feature-based, and relation-based knowledge distillation techniques in the offline scheme. The proposed knowledge distillation improves the mean squared error for RGB images from 2 to 0.14 weeks. The results are validated on three different datasets.