We investigate the compression of deep neural networks by quantizing their weights and activations into multiple binary bases, known as multi-bit (MBNs), which accelerate inference reduce storage for deployment on low-resource mobile embedded platforms. propose Adaptive Loss-aware Quantization (ALQ), a new MBN quantization pipeline that is able to achieve an average bitwidth below one-bit without notable loss in accuracy. Unlike previous solutions train quantizer minimizing error reconstruct...

Neuromorphic vision sensors are bio-inspired cameras that naturally capture the dynamics of a scene with unlit-low latency, filtering out redundant information low power consumption. Few works addressing object detection this sensor. In work, we propose to develop pedestrian detectors unlock potential event data by leveraging multi-cue and different fusion strategies. To make best data, introduce three event-stream encoding methods based on Frequency, Surface Active Event (SAE) Leaky...

With the surge of inexpensive computational and memory resources, neural networks (NNs) have experienced an unprecedented growth in architectural complexity. Introducing NNs to resource-constrained devices enables cost-efficient deployments, widespread availability, preservation sensitive data. This work addresses challenges bringing Machine Learning MCUs, where we focus on ubiquitous ARM Cortex-M architecture. The detailed effects trade-offs that optimization methods, software frameworks,...

Large Language Models (LLMs) have demonstrated remarkable capabilities in important tasks such as natural language understanding, generation, and complex reasoning the potential to make a substantial impact on our society. Such capabilities, however, come with considerable resources they demand, highlighting strong need develop effective techniques for addressing their efficiency challenges.In this survey, we provide systematic comprehensive review of efficient LLMs research. We organize...

Keypoint detector and descriptor are two main components of point cloud registration. Previous learning-based keypoint detectors rely on saliency estimation for each or farthest sample (FPS) candidate points selection, which inefficient not applicable in large scale scenes. This paper proposes Random Sample-based Detector Descriptor Network (RSKDD-Net) The key idea is using random sampling to efficiently select a method jointly generate keypoints descriptors. To tackle the information loss...

Vision Transformers (ViTs) have recently achieved promising results in various computer vision tasks. However, ViTs high computation costs and a large number of parameters due to the stacked multi-head self-attention (MHSA) expanded feed-forward network (FFN) modules. Since complexity Transformer-based models is quadratic with length input tokens, most current efforts focus on reducing tokens improve model efficiency. Unlike previous studies, we argue that diverse redundant features help...

Robotic grasping plays an important role in the field of robotics. The current state-of-the-art robotic detection systems are usually built on conventional vision, such as RGB-D camera. Compared to traditional frame-based computer neuromorphic vision is a small and young community research. Currently, there limited event-based datasets due troublesome annotation asynchronous event stream. Annotating large scale often takes lots computation resources, especially when it comes data for...

We design deep neural networks (DNNs) and corresponding networks' splittings to distribute DNNs' workload camera sensors a centralized aggregator on head mounted devices meet system performance targets in inference accuracy latency under the given hardware resource constraints. To achieve an optimal balance among computation, communication, performance, split-aware architecture search framework, SplitNets, is introduced conduct model designing, splitting, communication reduction...

For flow in open channels, rating curves are indicators by which readings of the depth converted to discharge. Rating have unique relations when is steady and uniform. unsteady flows, produced flood hydrographs, show characteristic loops, implying non-unique relations. In this paper, passage hydrographs investigated, focusing on basic hydraulic parameters, namely mean velocity, discharge depth, subsequently other important such as friction velocity coefficient. The sequence arrival their...

Data collected by IoT devices are often private and have a large diversity across users. Therefore, learning requires pre-training model with available representative data samples, deploying the pre-trained on devices, adapting deployed device local data. Such an on-device adaption for deep empowered applications demands memory efficiency. However, existing gradient-based meta schemes fail to support memory-efficient adaptation. To this end, we propose p-Meta, new method that enforces...

With the surge of inexpensive computational and memory resources, neural networks (NNs) have experienced an unprecedented growth in architectural complexity. Introducing NNs to resource-constrained devices enables cost-efficient deployments, widespread availability, preservation sensitive data. This work addresses challenges bringing Machine Learning MCUs, where we focus on ubiquitous ARM Cortex-M architecture. The detailed effects trade-offs that optimization methods, software frameworks,...

Deep neural networks (DNNs) have succeeded in many different perception tasks, e.g., computer vision, natural language processing, reinforcement learning, etc. The high-performed DNNs heavily rely on intensive resource consumption. For example, training a DNN requires high dynamic memory, large-scale dataset, and large number of computations (a long time); even inference with also demands amount static storage, time), energy. Therefore, state-of-the-art are often deployed cloud server...

Integrating Internet of things (IoT) techniques into automated vehicles has been a vision in intelligent transportation system, there is however seldom researches addressing it. To this end, we envision scenario: short‐range on‐board sensor perception system attached to individual mobile applications such as are connected via IoT and transferred long‐range mobile‐sensing which can be used part more extensive surveilling the environment. However, sensing brings new challenges for how...

The rapid growth of IoT era is shaping the future mobile services. Advanced communication technology enables a heterogeneous connectivity where devices broadcast information to everything. Mobile applications such as robotics and vehicles connecting cloud surroundings transfer short-range on-board sensor perception system long-range mobile-sensing system. However, sensing brings new challenges for how efficiently analyze intelligently interpret deluge data in mission- critical In this...

Robotic grasping plays an important role in the field of robotics. The current state-of-the-art robotic detection systems are usually built on conventional vision, such as RGB-D camera. Compared to traditional frame-based computer neuromorphic vision is a small and young community research. Currently, there limited event-based datasets due troublesome annotation asynchronous event stream. Annotating large scale dataset often takes lots computation resources, especially data for video-level...

We demonstrate optical memory and light-triggered electrical oscillations in a VO2 electro optic micro-wire device for potential applications neuromorphic computing architectures.

Emerging edge intelligence applications require the server to retrain and update deep neural networks deployed on remote nodes leverage newly collected data samples. Unfortunately, it may be impossible in practice continuously send fully updated weights these due highly constrained communication resource. In this paper, we propose weight-wise partial updating paradigm, which smartly selects a small subset of each server-to-edge round, while achieving similar performance compared full...

This paper describes an adaptive robust moving hands recognition algorithm using Kinect V2, which can detect the bare or with gloves in real-time image stream under complex lighting condition. Firstly, according to Bayes criterion, a novel skin color classification is built on best separation plane space, found through linear discriminant analysis (LDA). Secondly, learning rate and connected component theory are added traditional background subtraction. Finally, this new subtraction LDA...