Smart home hubs have significantly improved everyday home life by serving as central control units that connect and manage various devices, such as lights, door locks, curtains, and cameras. However, the limited CPU and memory resources of these hubs hinder the execution of multiple intelligent tasks, such as human activity recognition, image classification, and speech recognition. To fully utilize these resources, we proposeS-HUB, a scalable deep neural network (DNN) fusion framework for smart home hubs capable of handling multiple tasks. In the offline phase, we apply DNN pruning, weight virtualization, and re-fusion to create a unified model that dynamically scales across tasks. In the online phase, we design a DNN scheduling optimizer to achieve optimal multitask inference while adhering to resource constraints. Finally, comparative experiments and evaluations in smart home scenarios are conducted to assess the performance ofS-HUBacross various tasks, including speech recognition, object detection, gesture recognition, food recognition, and fall detection. Experimental results show that compared to two state-of-the-art baselines,S-HUBachieves an average task processing time of 3 seconds (at least 14% faster) under accuracy constraints, and an average accuracy loss rate of 4.21% (at least 68% lower) under latency constraints. In unconstrained scenarios, it consistently delivers the best overall performance (2.99 seconds and 2.94% loss), demonstrating the scalability and effectiveness of our fusion method in handling performance-resource trade-offs across tasks.

Multi-modal human activity recognition (HAR) is a key technology for a wide range of applications and has received widespread attention in recent years. However, the difficulty of achieving generalizability in multi-modal sensing models, combined with heterogeneous and unlabeled downstream data, significantly hinders their broader adoption. In this work, we proposePFHAR, a unified framework for practically adopting multi-modal foundation HAR model to target user groups.PFHARuses a novel dynamic masked contrastive learning method to pre-train a foundation model on various heterogeneous public HAR datasets, ensuring strong generalizability across different modal combinations. It then adopts semi-supervised edge-cloud collaborative learning to fine-tune the pre-trained model with heterogeneous and unlabeled local data, adapting it for the target user group. Our evaluations on public and self-collected datasets demonstrate thatPFHARsignificantly outperforms SOTA baselines in both the pre-training and edge-cloud collaborative fine-tuning stages.

Edge-cloud collaborative learning emerges as a promising paradigm for adapting pre-trained deep neural network (DNN) models to the ever-changing edge data environments and specific downstream tasks. However, the heterogeneity of edge devices and unlabeled local data hinder the effectiveness of existing collaborative learning approaches. To address the above issues, we proposeACL, a novel adaptive edge-cloud collaborative learning paradigm for heterogeneous devices with unlabeled local data. InACL, we first useFedNAS, a neural architecture search algorithm designed for collaborative learning to generate a customized model on each participating device, and then a lightweight semi-supervised collaborative learning frameworkHSSCLis used to fine-tune the pre-trained DNN model. Compared with the SOTA collaborative learning approaches,ACLachieves significant accuracy improvement, averaging 31.5% for image classification and 15.5% for object detection. Furthermore, it reduces time overhead by 3.1-5.1× and memory overhead by 6.3-12.5×. We will release our models and tools.

Deploying deep learning models to edge devices for low-latency and privacy-preserving applications has become a trend. To adapt to heterogeneous devices and data, it is significant to generate customized models. However, existing model adaptation approaches require edge devices to make interactions (collecting hardware information or local data) with the cloud, which raises privacy concerns, increases communication costs, and burdens the cloud. By contrast, we proposeCamoNet, a universal on-device model adaptation framework with zero interaction between devices and the cloud. InCamoNet, a lightweight on-device neural architecture search module is utilized to quickly generate a customized model for subsequent on-device training, followed by an on-device contrastive transfer learning module to effectively leverage unlabeled data for fine-tuning the customized model. Extensive experimental results show thatCamoNetcan effectively run on various edge devices. Compared with the SOTA model adaptation approaches,CamoNetachieves significant accuracy improvement by 25.2% on average for image classification, 10.1% on average for object detection, and reduces the training memory by 4.8-11.4×. We will open-source our models and tools for edge AI developers.