Recent advancements in on-device training for deep neural networks have underscored the critical need for efficient activation compression to overcome the memory constraints of mobile and edge devices. As activations dominate memory usage during training and are essential for gradient computation, compressing them without compromising accuracy remains a key research challenge. While existing methods for dynamic activation quantization promise theoretical memory savings, their practical deployment is impeded by system-level challenges such as computational overhead and memory fragmentation. To address these challenges, we introduce DAF, a Dynamic Activation Framework that enables scalable and efficient on-device training through system-level optimizations. DAF achieves both memory- and time-efficient dynamic quantization training by addressing key system bottlenecks. It develops hybrid reduction operations tailored to the memory hierarchies of mobile and edge SoCs, leverages collaborative CPU-GPU bit-packing for efficient dynamic quantization, and implements an importance-aware paging memory management scheme to reduce fragmentation and support dynamic memory adjustments. These optimizations collectively enable DAF to achieve substantial memory savings and speedup without compromising model training accuracy. Evaluations on various deep learning models across embedded and mobile platforms demonstrate up to a $22.9\times$ reduction in memory usage and a $3.2\times$ speedup, making DAF a scalable and practical solution for resource-constrained environments.

In this paper, an edge-side model training study is conducted on a resource-limited smart meter. The motivation of grid-edge intelligence and the concept of on-device training are introduced. Then, the technical preparation steps for on-device training are described. A case study on the task of photovoltaic power forecasting is presented, where two representative machine learning models are investigated: a gradient boosting tree model and a recurrent neural network model. To adapt to the resource-limited situation in the smart meter, "mixed"- and "reduced"-precision training schemes are also devised. Experiment results demonstrate the feasibility of economically achieving grid-edge intelligence via the existing advanced metering infrastructures.

Personal use of this material is permitted. Abstract --With the advent of Industry 5.0, manufacturers are increasingly prioritizing worker well-being alongside mass customization. Stress-aware Human-Robot Collaboration (HRC) plays a crucial role in this paradigm, where robots must adapt their behavior to human mental states to improve collaboration fluency and safety. This paper presents a novel framework that integrates Federated Learning (FL) to enable personalized mental state evaluation while preserving user privacy. By leveraging physiological signals, including EEG, ECG, EDA, EMG, and respiration, a multimodal model predicts an operator's stress level, facilitating real-time robot adaptation. The FL-based approach allows distributed on-device training, ensuring data confidentiality while improving model generalization and individual customization. Results demonstrate that the deployment of an FL approach results in a global model with performance in stress prediction accuracy comparable to a centralized training approach. Moreover, FL allows for enhancing personalization, thereby optimizing human-robot interaction in industrial settings, while preserving data privacy. The proposed framework advances privacy-preserving, adaptive robotics to enhance workforce wellbeing in smart manufacturing. With the rise of Industry 5.0, manufacturers are increasingly prioritizing worker well-being while addressing the growing demand for mass customization.

Out-of-Distribution (OOD) detection is critical for ensuring the reliability of machine learning models in safety-critical applications such as autonomous driving and medical diagnosis. While deploying personalized OOD detection directly on edge devices is desirable, it remains challenging due to large model sizes and the computational infeasibility of on-device training. Federated learning partially addresses this but still requires gradient computation and backpropagation, exceeding the capabilities of many edge devices. To overcome these challenges, we propose SecDOOD, a secure cloud-device collaboration framework for efficient on-device OOD detection without requiring device-side backpropagation. SecDOOD utilizes cloud resources for model training while ensuring user data privacy by retaining sensitive information on-device. Central to SecDOOD is a HyperNetwork-based personalized parameter generation module, which adapts cloud-trained models to device-specific distributions by dynamically generating local weight adjustments, effectively combining central and local information without local fine-tuning. Additionally, our dynamic feature sampling and encryption strategy selectively encrypts only the most informative feature channels, largely reducing encryption overhead without compromising detection performance. Extensive experiments across multiple datasets and OOD scenarios demonstrate that SecDOOD achieves performance comparable to fully fine-tuned models, enabling secure, efficient, and personalized OOD detection on resource-limited edge devices. To enhance accessibility and reproducibility, our code is publicly available at https://github.com/Dystopians/SecDOOD.

On-device training enables the model to adapt to new data collected from the sensors by fine-tuning a pre-trained model. Users can benefit from customized AI models without having to transfer the data to the cloud, protecting the privacy. However, the training memory consumption is prohibitive for IoT devices that have tiny memory resources. We propose an algorithm-system co-design framework to make on-device training possible with only 256KB of memory. On-device training faces two unique challenges: (1) the quantized graphs of neural networks are hard to optimize due to low bit-precision and the lack of normalization; (2) the limited hardware resource (memory and computation) does not allow full backpropagation.

On-device training of DNNs allows models to adapt and fine-tune to newly collected data or changing domains while deployed on microcontroller units (MCUs). However, DNN training is a resource-intensive task, making the implementation and execution of DNN training algorithms on MCUs challenging due to low processor speeds, constrained throughput, limited floating-point support, and memory constraints. In this work, we explore on-device training of DNNs for Cortex-M MCUs. We present a method that enables efficient training of DNNs completely in place on the MCU using fully quantized training (FQT) and dynamic partial gradient updates. We demonstrate the feasibility of our approach on multiple vision and time-series datasets and provide insights into the tradeoff between training accuracy, memory overhead, energy, and latency on real hardware.

Big Artificial Intelligence (AI) models have emerged as a crucial element in various intelligent applications at the edge, such as voice assistants in smart homes and autonomous robotics in smart factories. Training big AI models, e.g., for personalized fine-tuning and continual model refinement, poses significant challenges to edge devices due to the inherent conflict between limited computing resources and intensive workload associated with training. Despite the constraints of on-device training, traditional approaches usually resort to aggregating training data and sending it to a remote cloud for centralized training. Nevertheless, this approach is neither sustainable, which strains long-range backhaul transmission and energy-consuming datacenters, nor safely private, which shares users' raw data with remote infrastructures. To address these challenges, we alternatively observe that prevalent edge environments usually contain a diverse collection of trusted edge devices with untapped idle resources, which can be leveraged for edge training acceleration. Motivated by this, in this article, we propose collaborative edge training, a novel training mechanism that orchestrates a group of trusted edge devices as a resource pool for expedited, sustainable big AI model training at the edge. As an initial step, we present a comprehensive framework for building collaborative edge training systems and analyze in-depth its merits and sustainable scheduling choices following its workflow. To further investigate the impact of its parallelism design, we empirically study a case of four typical parallelisms from the perspective of energy demand with realistic testbeds. Finally, we discuss open challenges for sustainable collaborative edge training to point to future directions of edge-centric big AI model training.

On-device training is essential for neural networks (NNs) to continuously adapt to new online data, but can be time-consuming due to the device's limited computing power. To speed up on-device training, existing schemes select trainable NN portion offline or conduct unrecoverable selection at runtime, but the evolution of trainable NN portion is constrained and cannot adapt to the current need for training. Instead, runtime adaptation of on-device training should be fully elastic, i.e., every NN substructure can be freely removed from or added to the trainable NN portion at any time in training. In this paper, we present ElasticTrainer, a new technique that enforces such elasticity to achieve the required training speedup with the minimum NN accuracy loss. Experiment results show that ElasticTrainer achieves up to 3.5x more training speedup in wall-clock time and reduces energy consumption by 2x-3x more compared to the existing schemes, without noticeable accuracy loss.

On-device training has become an increasingly popular approach to machine learning, enabling models to be trained directly on mobile and edge devices. However, a major challenge in this area is the limited memory available on these devices, which can severely restrict the size and complexity of the models that can be trained. In this systematic survey, we aim to explore the current state-of-the-art techniques for breaking on-device training memory walls, focusing on methods that can enable larger and more complex models to be trained on resource-constrained devices. Specifically, we first analyze the key factors that contribute to the phenomenon of memory walls encountered during on-device training. Then, we present a comprehensive literature review of on-device training, which addresses the issue of memory limitations. Finally, we summarize on-device training and highlight the open problems for future research. By providing a comprehensive overview of these techniques and their effectiveness in breaking memory walls, we hope to help researchers and practitioners in this field navigate the rapidly evolving landscape of on-device training.

On-device training is essential for user personalisation and privacy. With the pervasiveness of IoT devices and microcontroller units (MCU), this task becomes more challenging due to the constrained memory and compute resources, and the limited availability of labelled user data. Nonetheless, prior works neglect the data scarcity issue, require excessively long training time (e.g. a few hours), or induce substantial accuracy loss ($\geq$10\%). We propose TinyTrain, an on-device training approach that drastically reduces training time by selectively updating parts of the model and explicitly coping with data scarcity. TinyTrain introduces a task-adaptive sparse-update method that dynamically selects the layer/channel based on a multi-objective criterion that jointly captures user data, the memory, and the compute capabilities of the target device, leading to high accuracy on unseen tasks with reduced computation and memory footprint. TinyTrain outperforms vanilla fine-tuning of the entire network by 3.6-5.0\% in accuracy, while reducing the backward-pass memory and computation cost by up to 2,286$\times$ and 7.68$\times$, respectively. Targeting broadly used real-world edge devices, TinyTrain achieves 9.5$\times$ faster and 3.5$\times$ more energy-efficient training over status-quo approaches, and 2.8$\times$ smaller memory footprint than SOTA approaches, while remaining within the 1 MB memory envelope of MCU-grade platforms.