Modern AI services must continually adapt to newly joined domains, yet delivering high-quality customized models is hampered by label sparsity, domain shifts, and tight budgets. We formulate this challenge as the learning system expansion problem and introduce HaT, an efficient heterogeneity-aware knowledge-transfer framework. HaT first selects a small set of high-quality source models with minimal overhead, and then fuses their imperfect predictions through a sample-wise attention mixer. Later, it adaptively distills the fused knowledge into target models via a knowledge dictionary. Extensive experiments on different tasks and modalities show that HaT outperforms state-of-the-art baselines by up to 16.5% accuracy, and saves 31.1% training time and up to 93.0% traffic.

The analog-to-digital converter (ADC) in a radio frequency (RF) front-end and digital signal processing (DSP) are significant sources of energy consumption, particularly in low-power systems like LoRa. To reliably demodulate weak signals, current systems rely on heavy oversampling - often 8x the signal bandwidth - which imposes a substantial and persistent oversampling tax on the analog front-end and DSP. This paper investigates if this tax can be mitigated by adapting techniques from image and video super resolution. We propose RF Super Resolution, a lightweight, real-time neural upscaler for RF signals. Our approach pairs an efficient digital interpolation algorithm with a shallow four-layer CNN. The neural network is trained to learn and correct the residual artifacts introduced by the digital upsampling and noise, effectively mimicking the output of a high-rate analog ADC and denoising filter. We validate our system on a large-scale, over-the-air LoRa study. Our results show that RF-SR, given a 2× Nyquist input (250 kHz), restores demodulation performance of a native 8× oversampled (2 MHz) system at half its sampling rate (1 MHz). This effectively removes the analog oversampling requirement, and provides an additional 1.25 dB SNR gain over the oversampled baseline, making it an efficient and effective signal enhancer suitable for gateway integration or post training quantized deployment at the end-node.

This paper presents the first nationwide deployment of human activity recognition (HAR) technology in the on-demand food delivery industry. We successfully adapted the state-of-the-art LIMU-BERT foundation model to the delivery platform. Spanning three phases over two years, the deployment progresses from a feasibility study in Yangzhou City to nationwide adoption involving 500,000 couriers across 367 cities in China. The adoption enables a series of downstream applications, and large-scale tests demonstrate its significant operational and economic benefits, showcasing the transformative potential of HAR technology in real-world applications. Additionally, we share lessons learned from this deployment and open-source our LIMU-BERT pretrained with millions of hours of sensor data.

Model splitting and offloading part of the DNN model from the mobile device to the cloud server, known as collaborative inference, improves end-to-end latency and server throughput in CNN-based models. However, current collaborative approaches do not apply to popular Transformer-based models, as these models generate large intermediate outputs with significant transmission latency and have a uniform block structure that makes it challenging to serve tail models efficiently on the server. We propose CollabTrans, a collaborative inference framework designed for Transformer-based models to enhance server scalability when handling requests from numerous end devices, taming the large output with truncated SVD and serving heterogeneous tail models by sharing a complete model. We demonstrate our framework with a real-time image classification mobile application together with background inference request traffic, showing the high inference throughput of our framework.

In an era where embedded and mobile devices are becoming ubiquitous, the intersection of Artificial Intelligence and the Internet of Things (AIoT) is rapidly transforming various fields. This paper delves into the challenges and innovations in this domain, particularly focusing on the development of generalized sensing models for wearable sensor data. We introduce novel approaches to leverage the abundant unlabeled sensor data, physical sensing knowledge, and common knowledge embedded in Large Language Models (LLMs) to enhance the generalizability of AIoT models significantly.

Due to the ubiquitous deployment of WiFi infrastructure, numerous studies have employed WiFi RSSI fingerprinting for indoor localization. However, fingerprinting methods necessitate labour-intensive site surveys for fingerprint collection and location annotation. To address these limitations, we propose WiMU, a real-time indoor localization system that integrates WiFi and inertial measurement unit (IMU) data to enhance the real-time performance and accuracy of localization. WiMU operates on commodity WiFi infrastructure without the need for additional hardware, leveraging crowd-sourced user trajectories to learn spatial representations of access points (APs). These representations can be fine-tuned with minimal labeled data to support effective localization. Extensive evaluations demonstrate that WiMU reduces the cost of building an indoor localization system while ensuring high positioning accuracy, paving the way for the large-scale deployment of real-time indoor localization systems.

Existing inertial measurement unit (IMU) based human activity recognition (HAR) approaches still face a major challenge when adopted across users in practice. The severe heterogeneity in IMU data significantly undermines model generalizability in wild adoption. This paper presents UniHAR, a universal HAR framework for mobile devices. To address the challenge of data heterogeneity, we thoroughly study augmenting data with the physics of the IMU sensing process and present a novel adoption of data augmentations for exploiting both unlabeled and labeled data. We consider two application scenarios of UniHAR, which can further integrate federated learning and adversarial training for improved generalization. UniHAR is fully prototyped on the mobile platform and introduces low overhead to mobile devices. Extensive experiments demonstrate its superior performance in adapting HAR models across four open datasets.

With deep learning, millimeter-wave radar-based gesture recognition applications have achieved satisfactory results. However, most existing approaches highly rely on highquality labeled data, and they suffer from severe over-fitting when labeled data are scarce. To end this, we present RadarAE, a novel representation learning framework for radar sensing applications. RadarAE learns sophisticated representations from massive low-cost unlabeled radar data, which enables accurate gesture recognition with few labeled data. To achieve this goal, we first meticulously observe the characteristics of raw radar data and extract an effective feature, Spatio-Temporal Motion Map (STMM). Then we borrow the key principle of Masked Autoencoders (MAE), a self-supervised learning technique for images, and propose an MAE-like model to learn useful representations from STMM. To adapt RadarAE to radar sensing applications, we present a series of customization techniques, including data augmentation, optimized model structure, and adaptive pretraining method. With the learned high-level representations, gesture recognition models can achieve superior performance in few-shot scenarios. Experiment results show that our model can achieve 79.1%, 92.1%, 97.8%, and 99.5% recognition accuracy in the 1, 2, 4, and 8-shot scenarios, respectively, where x-shot refers to the number of labeled samples for each gesture type. The source codes and dataset are made publicly available11https://githuh.com/Ela-Boska/RadarAE.

Deep learning greatly empowers Inertial Measurement Unit (IMU) sensors for various mobile sensing applications, including human activity recognition, human-computer interaction, localization and tracking, and many more. Most existing works require substantial amounts of well-curated labeled data to train IMU-based sensing models, which incurs high annotation and training costs. Compared with labeled data, unlabeled IMU data are abundant and easily accessible. In this work, we present LIMU-BERT, a novel representation learning model that can make use of unlabeled IMU data and extract generalized rather than task-specific features. LIMU-BERT adopts the principle of self-supervised training of the natural language model BERT to effectively capture temporal relations and feature distributions in IMU sensor measurements. However, the original BERT is not adaptive to mobile IMU data. By meticulously observing the characteristics of IMU sensors, we propose a series of techniques and accordingly adapt LIMU-BERT to IMU sensing tasks. The designed models are lightweight and easily deployable on mobile devices. With the representations learned via LIMU-BERT, task-specific models trained with limited labeled samples can achieve superior performances. We extensively evaluate LIMU-BERT with four open datasets. The results show that the LIMU-BERT enhanced models significantly outperform existing approaches in two typical IMU sensing applications.

In recent years, location-based services have been widely applied not only in daily life but also in automation industries. As one of main location sensing technologies, RFID based localization has attracted increasing attention. Existing synthetic aperture RFID localization systems use the inverse correlation filter to reconstruct holograms and achieve satisfactory accuracy. However, these methods require accurate aperture positions for theoretical signal construction, while the ubiquitous aperture uncertainty in practice causes non-negligible performance degradation. In this paper, we present PEC, an accurate synthetic aperture RFID localization system with aperture position error compensation, which has a major advantage over the classic systems for no need to know the exact trajectory of the synthetic aperture. We first build a mathematical model for localization and merge all coherent received signals to estimate the tag position. Then we propose an iterative algorithm which can alternately estimate both the tag position and the aperture position error. We have implemented and evaluated PEC using commercial-off-the-shelf (COTS) RFID devices. Extensive experimental results show that it achieves the cm-level accuracy with aperture position error in noisy environments, which proves its effectiveness and robustness.

In recent years, radio frequency identification (RFID)-based approaches have been demonstrated to be a promising indoor localization techniques for many valuable applications, such as tracking tagged objects on the manufacturing lines, locating items in smart warehouses, and so on. In the near future, many applications will gain great benefits from knowing the positions of RFID-tagged objects. However, existing localization approaches often suffer from severe accuracy degradation in real-world environments due to the prevalent environmental interferences, such as the multipath effects. To this end, we designed an RFID-based localization system FaHo, which leverages a deep learning enhanced holographic technique for locating RFID tags accurately even in complex indoor environments. By carefully analyzing the features of the traditional holographic method, we created a new hologram-based algorithm called joint hologram, which yields a robust likelihood for each assumed position to be the true tag position. FaHo then adopts a deep convolutional neural network for analyzing the whole hologram, and subsequently estimate the true location of the RFID tag rather than simply seek for the largest-likelihood location. Furthermore, we implemented FaHo and evaluated its performance in several multipath-rich scenarios. The experimental results show that FaHo can achieve centimeter-level accuracy in both the lateral and radial directions using only one moving antenna. More importantly, our work also demonstrates that hologram-based localization is a highly effective technique for RFID indoor localization tasks.

In the future, libraries and warehouses will gain benefits from the spatial location of books and merchandises attached with RFID tags. Existing localization algorithms, however, usually focus on improving positioning accuracy or the ordering one for RFID tags on the same layer. Nevertheless, books or merchandises are placed on the multilayer in reality and the layer of RFID tagged object is also an important position indication. To this end, we design PRMS, an RFID based localization system which utilizes both phase and RSSI values of the backscattered signal provided by a single antenna to estimate the spatial position for RFID tags. Our basic idea is to gain initial estimated locations of RFID tags through a basic model which extracts the phase differences between received signals to locate tags. Then an advanced model is proposed to improve the positioning accuracy combined with RF hologram based on basic model. We further change traditional deployment of a single antenna to distinguish the features of RFID tags on multilayer and adopt a machine learning algorithm to get the layer information of tagged objects. The experiment results show that the average accuracy of layer detection and sorting at low tag spacing ($2\sim8$cm) are about 93% and 84% respectively.