Continuously adapting pre-trained models to local data on resource constrained edge devices is the \emph{last mile} for model deployment. However, as models increase in size and depth, backpropagation requires a large amount of memory, which becomes prohibitive for edge devices. In addition, most existing low power neural processing engines (e.g., NPUs, DSPs, MCUs, etc.) are designed as fixed-point inference accelerators, without training capabilities. Forward gradients, solely based on directional derivatives computed from two forward calls, have been recently used for model training, with substantial savings in computation and memory. However, the performance of quantized training with fixed-point forward gradients remains unclear. In this paper, we investigate the feasibility of on-device training using fixed-point forward gradients, by conducting comprehensive experiments across a variety of deep learning benchmark tasks in both vision and audio domains. We propose a series of algorithm enhancements that further reduce the memory footprint, and the accuracy gap compared to backpropagation. An empirical study on how training with forward gradients navigates in the loss landscape is further explored. Our results demonstrate that on the last mile of model customization on edge devices, training with fixed-point forward gradients is a feasible and practical approach.

Look-up tables(LUTs)-based methods have recently shown enormous potential in image restoration tasks, which are capable of significantly accelerating the inference. However, the size of LUT exhibits exponential growth with the convolution kernel size, creating a storage bottleneck for its broader application on edge devices. Here, we address the storage explosion challenge to promote the capacity of mapping the complex CNN models by LUT. We introduce an innovative separable mapping strategy to achieve over $7\times$ storage reduction, transforming the storage from exponential dependence on kernel size to a linear relationship. Moreover, we design a dynamic discretization mechanism to decompose the activation and compress the quantization scale that further shrinks the LUT storage by $4.48\times$. As a result, the storage requirement of our proposed TinyLUT is around 4.1\% of MuLUT-SDY-X2 and amenable to on-chip cache, yielding competitive accuracy with over $5\times$ lower inference latency on Raspberry 4B than FSRCNN. Our proposed TinyLUT enables superior inference speed on edge devices with new state-of-the-art accuracy on both of image super-resolution and denoising, showcasing the potential of applying this method to various image restoration tasks at the edge.

Recently, a new trend of exploring sparsity for accelerating neural network training has emerged, embracing the paradigm of training on the edge. This paper proposes a novel Memory-Economic Sparse Training (MEST) framework targeting for accurate and fast execution on edge devices. The proposed MEST framework consists of enhancements by Elastic Mutation (EM) and Soft Memory Bound (&S) that ensure superior accuracy at high sparsity ratios. Different from the existing works for sparse training, this current work reveals the importance of sparsity schemes on the performance of sparse training in terms of accuracy as well as training speed on real edge devices. On top of that, the paper proposes to employ data efficiency for further acceleration of sparse training.

Scaling up the convolutional neural network (CNN) size (e.g., width, depth, etc.) is known to effectively improve model accuracy. However, the large model size impedes training on resource-constrained edge devices. For instance, federated learning (FL) may place undue burden on the compute capability of edge nodes, even though there is a strong practical need for FL due to its privacy and confidentiality properties. To address the resource-constrained reality of edge devices, we reformulate FL as a group knowledge transfer training algorithm, called FedGKT. FedGKT designs a variant of the alternating minimization approach to train small CNNs on edge nodes and periodically transfer their knowledge by knowledge distillation to a large server-side CNN.

The exponential growth of Internet-connected devices has presented challenges to traditional centralized computing systems due to latency and bandwidth limitations. Edge computing has evolved to address these difficulties by bringing computations closer to the data source. Additionally, traditional machine learning algorithms are not suitable for edge-computing systems, where data usually arrives in a dynamic and continual way. However, incremental learning offers a good solution for these settings. We introduce a new approach that applies the incremental learning philosophy within an edge-computing scenario for the industrial sector with a specific purpose: real time quality control in a manufacturing system. Applying continual learning we reduce the impact of catastrophic forgetting and provide an efficient and effective solution.

Video Understanding, Scene Interpretation and Commonsense Reasoning are highly challenging tasks enabling the interpretation of visual information, allowing agents to perceive, interact with and make rational decisions in its environment. Large Language Models (LLMs) and Visual Language Models (VLMs) have shown remarkable advancements in these areas in recent years, enabling domain-specific applications as well as zero-shot open vocabulary tasks, combining multiple domains. However, the required computational complexity poses challenges for their application on edge devices and in the context of Mobile Robotics, especially considering the trade-off between accuracy and inference time. In this paper, we investigate the capabilities of state-of-the-art VLMs for the task of Scene Interpretation and Action Recognition, with special regard to small VLMs capable of being deployed to edge devices in the context of Mobile Robotics. The proposed pipeline is evaluated on a diverse dataset consisting of various real-world cityscape, on-campus and indoor scenarios. The experimental evaluation discusses the potential of these small models on edge devices, with particular emphasis on challenges, weaknesses, inherent model biases and the application of the gained information. Supplementary material is provided via the following repository: https://datahub.rz.rptu.de/hstr-csrl-public/

--Processing-in-Memory (PIM) architectures offer promising solutions for efficiently handling AI applications in energy-constrained edge environments. While traditional PIM designs enhance performance and energy efficiency by reducing data movement between memory and processing units, they are limited in edge devices due to continuous power demands and the storage requirements of large neural network weights in SRAM and DRAM. Hybrid PIM architectures, incorporating nonvolatile memories like MRAM and ReRAM, mitigate these limitations but struggle with a mismatch between fixed computing resources and dynamically changing inference workloads. T o address these challenges, this study introduces a Heterogeneous-Hybrid PIM ( HH-PIM) architecture, comprising high-performance MRAM-SRAM PIM modules and low-power MRAM-SRAM PIM modules. We further propose a data placement optimization algorithm that dynamically allocates data based on computational demand, maximizing energy efficiency. FPGA prototyping and power simulations with processors featuring HH-PIM and other PIM types demonstrate that the proposed HH-PIM achieves up to 60.43% average energy savings over conventional PIMs while meeting application latency requirements. These results confirm HH-PIM's suitability for adaptive, energy-efficient AI processing in edge devices. With the advent of artificial intelligence (AI), real-world applications are rapidly expanding, fueling a trend to embed AI capabilities into IoT devices across diverse fields. However, traditional server-centric data processing, such as cloud computing, faces significant energy and latency challenges due to processing and communication overloads.

Large Multimodal Models (LMMs) are inherently modular, consisting of vision and audio encoders, projectors, and large language models. Yet, they are almost always executed monolithically, which underutilizes the heterogeneous accelerators (NPUs, GPUs, DSPs) in modern SoCs and leads to high end-to-end latency. In this paper, we present NANOMIND, a hardware--software co-design inference framework for Large Multimodal Models (LMMs) that breaks large models into modular ``bricks'' (vision, language, audio, etc.) and maps each to its ideal accelerator. The key insight is that large models can be broken into modular components and scheduled to run on the most appropriate compute units. It performs module-level dynamic offloading across accelerators on unified-memory SoCs. By combining customized hardware design, system-level scheduling, and optimized low-bit computation kernels, we demonstrate our framework with a compact, battery-powered device capable of running LMMs entirely on device. This prototype functions as a self-contained intelligent assistant that requires no network connectivity, while achieving higher throughput and superior power efficiency under strict resource constraints. The design further bypasses CPU bottlenecks and reduces redundant memory usage through token-aware buffer management and module-level coordination. Our system outperforms existing implementations in resource efficiency, cutting energy consumption by 42.3\% and GPU memory usage by 11.2\%. This enables a battery-powered device to run LLaVA-OneVision with a camera for nearly 20.8 hours.

Large Language Models (LLMs) demonstrate impressive capabilities in natural language understanding and generation, but incur high communication overhead and privacy risks in cloud deployments, while facing compute and memory constraints when confined to edge devices. Cloud-edge inference has emerged as a promising paradigm for improving privacy in LLM services by retaining sensitive computations on local devices. However, existing cloud-edge inference approaches apply uniform privacy protection without considering input sensitivity, resulting in unnecessary perturbation and degraded utility even for non-sensitive tokens. To address this limitation, we propose Privacy-aware Routing for Inference with Semantic Modulation (PRISM), a context-aware framework that dynamically balances privacy and inference quality. PRISM executes in four stages: (1) the edge device profiles entity-level sensitivity; (2) a soft gating module, also on the edge, selects an execution mode -cloud, edge, or collaboration; (3) for collaborative paths, the edge applies adaptive two-layer local differential privacy based on entity risks; and (4) the cloud LLM generates a semantic sketch from the perturbed prompt, which is then refined by the edge-side small language model (SLM) using local context. Our results show that PRISM consistently achieves superior privacy-utility trade-offs in various scenarios, reducing energy consumption and latency to 40-50% of baseline methods such as Uniform and Selective LDP, while maintaining high output quality under strong privacy constraints. These findings are validated through comprehensive evaluations involving realistic prompts, actual energy measurements, and heterogeneous cloud-edge model deployments.