on-device-ai

Image super-resolution (SR) is widely used on mobile devices to enhance user experience. However, neural networks used for SR are computationally expensive, posing challenges for mobile devices with limited computing power. A viable solution is to use heterogeneous processors on mobile devices, especially the specialized hardware AI accelerators, but the reduced arithmetic precision on AI accelerators can lead to degraded perceptual quality in upscaled images. To address this limitation, we present a novel image SR technique that enhances the perceptual quality of upscaled images when using heterogeneous processors for SR computations. It strategically splits the SR model and dispatches different layers to heterogeneous processors, to meet the time constraint while minimizing the impact of AI accelerators on image quality. Experiment results show that our method outperforms the best baselines, improving perceptual image quality by up to 2×, or reducing SR computing latency by up to 5.6× with on-par image quality.

Being different from model compression that requires expensive retraining, sparse activation can effectively reduce neural network models’ inference cost at runtime without any prior retraining or adaptation efforts. Although sparse activation has been proved to be effective on Large Language Models (LLMs) that are usually redundant (e.g., OPT and BLOOMZ models), its applicability on recent Small Language Models (SLMs) with higher parameter efficiency remains questionable. Our recent work verified such possibility by using gradient-based attribution scores to evaluate neurons’ importance in inference, in both analytical and experimental perspectives. Our results show that we can achieve up to 80% sparsity in major SLM models, including Phi-1.5/2 and MobiLlama-0.5B/1B, with less than 5% model accuracy loss on QA tasks.

This is the first work that allows multimodal LLMs to elastically switch between input data modalities at runtime, for embodied AI applications such as autonomous navigation. Our basic technical approach is to use fully trainable projectors to adaptively connect the unimodal data encoders being used to a flexible set of last LLM blocks. In this way, we can flexibly adjust the amount of LLM blocks being connected to balance between accuracy of runtime fine-tuning cost, and optimize the efficiency of cross-modal interaction by controlling the amount of information being injected in each connection. Our implementations on NVidia Jetson AGX Orin demonstrate short modality adaptation delays of few minutes with mainstream LLMs, 3.7x fine-tuning FLOPs reduction, and 4% accuracy improvements on multimodal QA tasks.

The growing need of fine-tuning large language models (LLMs) can lead to significant energy consumption and environmental impact. To address this issue, we introduce GreenTrainer, a novel LLM fine-tuning technique. GreenTrainer assesses the backpropagation costs and contributions of different tensors to model accuracy, allowing for the selection of the most efficient set of tensors. This selection is guided by a user-defined objective, which can adapt to energy supply considerations and Green AI goals. Experimental results demonstrate that GreenTrainer can reduce FLOPs by up to 64% without compromising model accuracy, and outperforms existing techniques like LoRA while maintaining comparable FLOPs reduction.

Federated Learning (FL) efficiency is influenced by intertwined data and device heterogeneities. Traditionally, these factors are treated separately, which becomes ineffective in addressing staleness issue due to asynchronous FL. We introduce a novel FL framework employing the gradient inversion technique to get estimations of clients’ local training data from their uploaded stale model updates, and use these estimations to compute non-stale client model updates, which addresses both data quality and privacy concerns. Experiments on mainstream datasets reveal our approach enhances model accuracy by up to 20% and accelerates FL training by up to 35% over existing methods.

The first on-device AI technique that achieves full elasticity of on-device training on resource-constrained mobile and embedded devices. By leveraging the principle of eXplainable AI (XAI) and evaluating the importance of different tensors in training, we allow fully flexible adaptation of the trainable neural network portion at runtime, according to the current training needs and online data patterns, to minimize the training cost without accuracy loss.

This work applies on-device AI techniques to interference cancellation in WiFi networks and enables generalizable interference cancellation on commodity WiFi devices without any extra RF hardware. By using neural network models to mimic WiFi network’s PHY-layer operation, AiFi can be generally applied to different types of interference signals ranging from concurrent WiFi transmissions, ZigBee/Bluetooth to wireless baby monitors or even microwave oven, and improves the MAC-layer frame reception rate by 18x.

AgileNN is the first work that achieves real-time inference (<20ms) of mainstream neural network models (e.g., ImageNet) on extremely weak MCUs (e.g., STM32 series with <1MB of memory), without impairing the inference accuracy. The usage of eXplainable AI (XAI) techniques allows >6x improvement of feature compressibility during offloading and >8x reduction of the local device’s resource consumption.