DeepSense

Huynh, Loc N.; Balan, Rajesh Krishna; Lee, Mi Kyung

doi:10.1145/2935643.2935650

Cited by 47 publications

(5 citation statements)

References 5 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…Mobile GPUs however are becoming more powerful, driven in part by machine vision and deep neural network applications (e.g. DeepSense [20]). Beyond efficiency, we can also exploit the sum of the computing power available across two machines.…”

Section: Fig 1: Logical Flow Diagram For Prototypical Implementationmentioning

confidence: 99%

Quality of Service Impact on Edge Physics Simulations for VR

Friston

Griffith

Swapp

et al. 2021

IEEE Trans. Visual. Comput. Graphics

View full text Add to dashboard Cite

Mobile HMDs must sacrifice compute performance to achieve ergonomic and power requirements for extended use. Consequently, applications must either reduce rendering and simulation complexity -along with the richness of the experience -or offload complexity to a server. Within the context of edge-computing, a popular way to do this is through render streaming. Render streaming has been demonstrated for desktops and consoles. It has also been explored for HMDs. However, the latency requirements of head tracking make this application much more challenging. While mobile GPUs are not yet as capable as their desktop counterparts, we note that they are becoming more powerful and efficient. With the hard requirements of VR, it is worth continuing to investigate what schemes could optimally balance load, latency and quality. We propose an alternative we call edge-physics: streaming at the scene-graph level from a simulation running on edge-resources, analogous to cluster rendering. Scene streaming is not only straightforward, but compute and bandwidth efficient. The most demanding loops run locally. Jobs that hit the power-wall of mobile CPUs are off-loaded, while improving GPUs are leveraged, maximising compute utilisation. In this paper we create a prototypical implementation and evaluate its potential in terms of fidelity, bandwidth and performance. We show that an effective system which maintains high consistencies on typical edge-links can be easily built, but that some traditional concepts are not applicable, and a better understanding of the perception of motion is required to evaluate such a system comprehensively.

show abstract

Section: Fig 1: Logical Flow Diagram For Prototypical Implementationmentioning

confidence: 99%

Quality of Service Impact on Edge Physics Simulations for VR

Friston

Griffith

Swapp

et al. 2021

IEEE Trans. Visual. Comput. Graphics

View full text Add to dashboard Cite

show abstract

“…CNN implementation using CUDA cuDNN [86] Nvidia's CUDA SDK for CNN implementation OpenCV DNN [87] OpenCV library with DNN and optimized for GPU Tensorflow [88] CNN framework with GPU (CUDA) performance optimization PyTorch [89] Optimized tensor library for DL using GPU (CUDA) DeepSense [90] Mobile GPU-based CNN optimization using OpenCL…”

Section: Cuda-convnet2 [85]mentioning

confidence: 99%

Best Practices for the Deployment of Edge Inference: The Conclusions to Start Designing

2021

View full text Add to dashboard Cite

The number of Artificial Intelligence (AI) and Machine Learning (ML) designs is rapidly increasing and certain concerns are raised on how to start an AI design for edge systems, what are the steps to follow and what are the critical pieces towards the most optimal performance. The complete development flow undergoes two distinct phases; training and inference. During training, all the weights are calculated through optimization and back propagation of the network. The training phase is executed with the use of 32-bit floating point arithmetic as this is the convenient format for GPU platforms. The inference phase on the other hand, uses a trained network with new data. The sensitive optimization and back propagation phases are removed and forward propagation is only used. A much lower bit-width and fixed point arithmetic is used aiming a good result with reduced footprint and power consumption. This study follows the survey based process and it is aimed to provide answers such as to clarify all AI edge hardware design aspects from the concept to the final implementation and evaluation. The technology as frameworks and procedures are presented to the order of execution for a complete design cycle with guaranteed success.

show abstract

“…Due to this model complexity, several studies have proposed the need for modifying CNN, through a combination with other approaches, in order to solve the computation cost problem [1] [2] [3], or dividing the burden, using another resources [4] [27] [26]. In addition, mobile computation context adopted in prior studies tend to address this issue by utilizing the Mobile GPU, in an attempt to minimize computation latency [27]. However, rather than place the entire load and process on a mobile device, it was proposed that another alternative, which exploits the capability of mobile internet service by introducing mobile client-server architecture.…”

Section: Fig 2 Pooling Layermentioning

confidence: 99%

“…The concept of the convolution neural network Convolution Neural Network (CNN) is a model that is expanded based on a traditional neural network, which is widely used for image detection and recognition [27] [17], specifically developed to process two-dimensional data. However, its weakness is identified in the complexity of the constructed number of layers [28], which has implications for the time of data training, hence, numerous studies recommend the use of additional computation resources, e.g., GPU (Graphics Processing Unit), in order to avoid suffering training cost [29] [27].…”

mentioning

confidence: 99%

The Javanese Letters Classifier with Mobile Client-Server Architecture and Convolution Neural Network Method

Harjoseputro

Handarkho

Adie

2019

Int. J. Interact. Mob. Technol.

View full text Add to dashboard Cite

<p class="0abstract">the rapid development of mobile technologies allows platform devices to perform sophisticated tasks, including character recognition. These identification systems are notable techniques that required high computation cost, in order to achieve acceptable accuracy resulting from diversity in alphabet shape and method of writing, especially for the non-Latin alphabet, e.g., Javanese letter. In addition, numerous studies have attempted to address these issues by employing a Convolution Neural Network (CNN) due to its ability to provide high accuracy in character detection. However, the performance on mobile devices is possibly faced with problems resulting from the limitation of computation resource on the platform that also affect computation cost. This study, therefore, proposes a 2-tier architecture by placing the mobile app as a client that invokes a Javanese letters classifier service, which is based on CNN, and implemented in the web-server through the Application Program Interface (API). The results show that the letter classification was successfully implemented in a mobile platform, with an accuracy rate of 86.68%, utilizing training for 50 epochs, and an average time of 1935 ms.</p>

show abstract

DeepSense

Cited by 47 publications

References 5 publications

Quality of Service Impact on Edge Physics Simulations for VR

Quality of Service Impact on Edge Physics Simulations for VR

Best Practices for the Deployment of Edge Inference: The Conclusions to Start Designing

The Javanese Letters Classifier with Mobile Client-Server Architecture and Convolution Neural Network Method

Contact Info

Product

Resources

About