Deep Learning used in Mobile Devices : An Overview

D.R.Patil

Research Scholar, Suresh Gyan Vihar University, Jaipur, Rajastan

patil.59712@mygyanvihar.com

Dr. Rajesh Purohit

Professor, Suresh Gyan Vihar University, Jaipur, Rajastan

gvset@mygyanvihar.com

 

Abstract :-

Numerous applications frequently embrace deep learning calculations, for example, CNN and RNN to separate data from mobile devices. Without being influenced the high exactness, utilization of deep learning calculations in mobile devices raises basic difficulties, i.e., high handling strength and power utilization. In This paper gives review on various programming structures for mobile devices created in deep learning. We principally center around segments of deep learning for portable situations like programming system that are enhanced for versatile conditions, low multifaceted nature calculation, particular equipment that is a piece of mobile devices for supporting the computationally costly procedures of deep system preparing and induction.

 

Keywords :- Deep Learning, Mobile Computing

 

Introduction :-

Today deep learning are most precise and factual model to frame a feeling or choosing that something is valid from the data originated from regular client practices and setting based on algorithms[1] – a creative territory of AI that is quickly changing how uproarious complex information from this present reality is demonstrated. The scope of derivation undertakings affected by deep learning incorporates the acknowledgment of: faces [2], feelings [3], objects [4] and words [5]. Anyway such deductions are basic to numerous versatile applications not many of them have embraced deep learning procedures. One of the incredible sensors of human conduct and setting are mouthpieces. Advances in sound based computational models of conduct and setting keep on widening the scope of deductions accessible to portable clients [6]. Through the amplifier it is conceivable to induce, for instance: every day exercises (e.g., eating [7], hacking [8], driving [9]), interior client states (e.g.,stress [10], feeling [11]) and encompassing conditions (e.g., number of close-by individuals [12]). In tale versatile applications sound detecting has advanced into a key structure obstruct that empower clients to screen and improve their wellbeing and prosperity [13], efficiency [14] and condition [15].

On versatile processor execution of deep learning calculation, for example, convolutional neural system is non-paltry because of serious computational prerequisites. DeepSense – a versatile GPU-based deep convolutional neural system (CNN) structure. For its plan, it investigated the contrasts between server-class and portable class GPUs, and considered adequacy of different streamlining methodologies. DeepSense[16] system, to execute deep learning calculations on mobile devices without cloud offloading. By utilizing versatile graphical preparing unit (GPU) as of late coordinated into mobile devices, the point is to help designers for 1) embracing a wide scope of existing DNN, CNN models prepared to keep running on server-class machines with negligible programming exertion, 2) accomplishing constant or delicate ongoing inertness for persistent detecting and intercession, 3) limiting vitality utilization on the registering mobile devices. DeepSense structure is developed on OpenCL [17], which is currently authoritatively upheld by various versatile GPUs, for example, Adreno and Mali.

For camera-prepared mobile devices, glimpse is a ceaseless, ongoing item acknowledgment framework. Impression catches full-movement video, finds objects of intrigue, perceives and names them, and tracks them from casing to outline for the client. Since the calculations for article acknowledgment involve critical calculation, Impression runs them on server machines. At the point when the inactivity between the server and mobile devices is higher than a casing time, this methodology brings down item acknowledgment precision. To recapture precision, Impression utilizes a functioning reserve of video outlines on the versatile device[18].

MCDNN considers applying model improvement to the multiprogrammed, spilling setting. various applications will look to run different DNNs on approaching high-datarate sensor streams, for example, sound, profundity, video and warm video. The enormous number of synchronous DNNs in activity and the high recurrence of their utilization will strain accessible assets, notwithstanding when streamlined models are utilized. to address this issue. I) model enhancement regularly permits an effortless tradeoff among precision and asset use. In this manner, a framework could adjust to high remaining burdens by utilizing less precise variations of upgraded models. ii) both spilling and multi-programming themselves give extra structure that empower ground-breaking new model optimizations[19].

DeepX [20] then empowered the execution of deep learning on mobile devices by part calculations over different processors.

DeepShark shares an abnormal state idea with DeepMon [21] (about re-utilizing the calculation asset), DeepShark’s reserving procedure eases the middle incomplete outcomes in the wake of preparing convolutional layers, empowering substantially more fine-grained sharing of calculation asset over the whole mobile device.

 

  1. Literature Survey:

 

2.1 Lane et al. stepped toward the constant execution of deep learning on mobile device [20].

            2.1.1 Deep Neural Network : group a sensor information is performed with a DNN utilizing a feed-forward calculation that works on each fragment of information independently. The calculation begin at the info layer and continuously pushes ahead layer by layer. At each layer feed-forward updates the condition of every unit individually. This procedure ends once all units in the yield layer are refreshed. The deduced class relates to the yield layer unit with the biggest state[22].

            2.1.2 Convolutional Neural Network : CNNs an elective plan of deep learning. Essentially CNNs utilized for vision and picture related tasks[23], yet their utilization is growing. A CNN is frequently made of at least one convolutional layers, pooling or sub-inspecting layers, and completely associated layers. The essential thought in CNN models is to extricate straightforward highlights from the high goals information picture and afterward changing over them into progressively complex highlights at much low goals at the higher layers. This is accomplished by :- a) applying different convolutional channels to catch nearby information properties. b) pursue max/min pooling layers causing separated highlights to be invariant to interpretations, this additionally goes about as a type of dimensionality decrease.

            2.1.3 Mobile Sensing Apps : The normal component between portable detecting applications is they all include the accumulation and comprehension of sensor information. To succeed this they insert AI calculations into their application structures. For these portable applications DeepX is intended to be utilized as a black-box by designers and give a substitution induction execution condition for any deep learning model they embrace. Sensor applications that persistently decipher information present the most testing situation as they may perform derivation on various occasions a moment; and in this way per-surmising asset utilization must be little if the application is to have great battery life. Applications that sense less consistently then again can bear the cost of higher per deduction costs. In any case, deep models need asset streamlining before they can even execute on a mobile platform [24]; numerous deep models have memory necessities that are unreasonably high for a versatile SoC to help. Essentially, execution times can without much of a stretch surpass limits that are acknowledged to an application (e.g., 30 seconds), showing an issue regardless of whether the surmising is sporadically actuated by the client for the duration of the day. One potential arrangement is runtime pressure of completely associated deep design layers to lessen memory prerequisites and surmising times.

            2.1.4 New Processors Developing on Portable SoCs : As the SoCs in mobile devices advance they are crushing in an undeniably wide scope of various computational units (GPUs, low-control CPU centers, multi-center CPUs). Indeed, even the Android-based LG G Watch R [25] incorporates a Snapdragon 400 [26] that contains a blending of DSP and a double center CPU. When performing various sorts of calculation every processor exhibits its very own asset profile. Contingent upon layer type or other trademark it makes distinctive exchange offs for them to execute segments of a deep model design, this assorted variety is generally new for mobile devices and utilized a layer-wise apportioning approach pursued by unraveling an enhancement condition.

 

2.2 DeepEar [27] demonstrated the plausibility of running whole DNNs. DeepEar can play out a master sound detecting framework for each undertaking.

            2.2.1 EmotionSense : As definite in[28] The list of capabilities traverses 32 PLP descriptors (16 static and 16 deltas) and 128-segment widespread (i.e., one model for all conditions) GMMs give characterization. Note, [29] depicts a speaker distinguishing proof classifier however they just receive the feeling acknowledgment pipeline.

            2.2.2 StressSense[29] determines 19 MFCCs alongside 7 other hand-picked highlights. To recreate the general model proposed in [29], they execute each of the 26 includes and give them to 16-part GMMs. Be that as it may, they don’t test the versatile variant of the framework.

            2.2.3 SpeakerSense[30] additionally utilizes 19 MFCCs. 32-part GMMs give order, and are prepared with a change restricting system [31] that brings down the effect of loud information. A 5s. smoothing window is connected to classifier results. The sound pipeline of [32] utilizes a 13-measurement MFCC vector joined by 4 extra unearthly highlights. Surmising happens with 32-segment GMMs and a 384 ms. sliding window to smooth results[32].

 

2.3 DeepSense [16] introduced early proof that utilizing a GPU could help improve the idleness of deep learning calculations. Two fundamental procedures of DeepSense are OpenCL and CNN.

            2.3.1 OpenCL : is a structure to help parallel programming crosswise over heterogeneous stages, for example, CPUs, GPUs or even DSPs. As of late, OpenCL has been broadly bolstered on both famous mobile device processors (e.g., Snapdragon and Exynos) and well known portable GPUs(e.g., Adreno and Mali)[17].

            2.3.2 Convolutional Neural System : Convolutional neural system (CNN) is a kind of feedforward neural system that is broadly received for picture and video acknowledgment [33,34].

DeepShark expands that work by giving a lot more improvements, a full execution, and broad assessment.

 

2.4 Glimpse [18] use the cloud to empower ongoing item recognition and following Cameras of good quality are presently accessible on pretty much every handheld and wearable mobile device. The high goals of these cameras combined with inescapable remote availability makes it possible to create consistent, constant article acknowledgment applications.

            2.4.1 Object Recognition Pipeline :

The acknowledgment pipeline for articles or faces, which keeps running on a video outline, comprises of three phases: location, include extraction, and acknowledgment.

            2.4.2 Detection: This stage scans the edge for objects of intrigue and finds them with bouncing boxes, yet without names. The article identifier utilizes particular attributes of items, for example, a shut limit for vehicles, a noticeable shading contrast from the foundation for street signs, and district contrasts for appearances (the eyes are darker than the cheeks).

            2.4.3 Feature extraction: This stage forms the substance inside the bouncing box registered by the finder to concentrate includes that speak to the article, utilizing strategies like the SIFT[35] and SURF[36].

            2.4.4 Acknowledgment/Marking: This stage perceives the article and doles out a name to it utilizing an AI classifier prepared disconnected utilizing a database of named item pictures. The preparation stage concentrates include vectors as referenced above, after which it builds a model, for example, a Help Vector Machine [37].

 

2.5 DeepX[20] : The structure and usage of DeepX, a programming quickening agent for profound learning execution. DeepX significantly brings down the gadget assets (viz. memory, calculation, vitality) required by deep learning that at present go about as a serious bottleneck to versatile selection. The establishment of DeepX is a pair of asset control calculations, intended for the deduction phase of profound realizing, that: (1) break down solid deep model system structures into unit-squares of different kinds, that are then more productively executed by heterogeneous neighborhood gadget processors (e.g., GPUs, CPUs); and (2), perform principled asset scaling that alters the design of deep models to shape the overhead every unit-squares presents. Investigations appear, DeepX can permit even enormous scale deep learning models to execute proficiently on present day versatile processors and essentially beat existing arrangements, for example, cloud-based offloading.

 

2.6 DeepMon[21] : The quick development of head-mounted gadgets, for example, the Microsoft Holo-focal point empowers a wide assortment of persistent vision applications. Such applications frequently receive profound learning calculations for example, CNN and RNN to remove rich relevant data from  the primary individual view video streams. In spite of the high precision, use of deep learning calculations in mobile devices raises basic challenges, i.e., high handling inactivity and power utilization. DeepMon, a versatile deep learning induction system to run an assortment of deep learning surmisings simply on a portable gadget in a quick and vitality effective way. For this the improvement methods to effectively offload convolutional layers to portable GPUs and quicken the handling; note that the convolutional layers are the normal execution bottleneck of numerous deep learning models. The test results appear that DeepMon can characterize a picture over the VGG-VeryDeep-16 deep learning model in 644ms on Samsung World S7, taking an significant advance towards ceaseless vision without forcing any privacy concerns nor systems administration cost[21].

 

2.7 MCDNN venture [19] executed deep learning calculations crosswise over mobile devices and mists. MCDNN gives shared, proficient framework for portable cloud applications that need to process gushing information (particularly video) utilizing Deep Neural Systems (DNNs). The AI people group has concentrated on decreasing the overhead of DNN execution. For significant models, for example, discourse and item acknowledgment, inquire about groups have spent impressive exertion delivering physically streamlined renditions of individual DNNs that are productive at run-time [38,39]. A few late endeavors in the AI people group have acquainted mechanized strategies with advance DNNs, for the most part variations of grid factorization and sparsification to lessen space [40,41] and computational requests [42]. A significant number of these endeavors bolster the people astuteness that for asset use DNN precision can extensively be exchanged off. MCDNN is supporting to these endeavors, in that it is freethinker to the specific enhancements used to create model variations that exchange off execution precision for assets. The notoriety of head-mounted augmented reality (AR) gadgets, for example, the Microsoft Hololens [44] and the Google Glass [43] has offered ascend to another class of constant versatile vision applications. These range from distinguishing street signs progressively to give headings [45], to recognizing individuals in the earth to offer direction to people experiencing dementia [46]. In all these utilization cases, the shared trait is the need to perform PC vision calculations progressively on a persistent video stream given by the AR gadgets.

 

2.8 DeepShark[47] : Deep learning gives new chances to versatile applications to accomplish higher execution than previously. Or maybe, the deep learning execution on mobile phones today is to a great extent requesting on costly asset overheads, forces a huge load on the battery life and constrained memory space. Existing strategies either use cloud or edge foundation that require to transfer client information, be that as it may, bringing about a danger of protection spillage and enormous information moves; or receive packed deep models, by the by, minimizing the calculation exactness. In DeepShark, a stage to empower mobile phones with the capacity of adaptable asset assignment in utilizing business off-the-rack (Bunks) deep learning frameworks. Contrasted with existing methodologies, DeepShark looks for a fair point among time and memory productivity by client necessities, separates modern deep model into code square stream and steadily executes such squares on framework on-chip (SoC). Hence, DeepShark requires fundamentally less memory space on mobile device and accomplishes the default exactness. What’s more, all alluded client information of model preparing is dealt with locally, consequently to maintain a strategic distance from superfluous information move and system inertness. DeepShark is presently created on two bunks deep learning frameworks, i.e., Caffe and TensorFlow. The exploratory assessments show its adequacy in the parts of memory space and vitality cost.

 

Conclusion :

In this paper, we examine the DeepEar, DeepMon, DeepX, Deepsense, MCDNN, Glimps, DeepShark used in mobile devices. For sound detecting applications on low-control versatile DSPs DeepEar demonstrated the attainability of running whole DNNs. DeepX then empowered the execution of DNN and CNN on mobile devices by part calculations over various co-processors. DeepMon can centers around lessening the inactivity of convolutional layers. Additionally, it bolsters OpenCL, Vulkan and GPUs on monetarily accessible mobile devices in the market. DeepSense displayed that utilizing a GPU could help improve the inactivity of DNN calculations – DeepMon expands that work by giving a lot more advancements, a full usage, and broad assessment. Impression utilized the cloud to empower ongoing article identification and following while MCDNN executed deep learning calculations crosswise over mobile devices and mists. Moreover, while DeepMon shares an abnormal state idea with MCDNN (about re-utilizing the DNN calculation), DeepMon’s storing system re-utilizes the middle fractional outcomes while preparing convolutional layers, empowering substantially more fine-granule sharing of calculation crosswise over back to back pictures. DeepShark, a novel stage which permits adaptable asset designation in Beds versatile deep learning frameworks.

References :

  1. Y. Bengio, et al., “Deep learning,” MIT Press 2015.
  2. Y. Taigman, et al., “Deepface: Closing the gap to human-level performance in face verification,” CVPR ’14
  3. Y. Kim, et al., “Deep learning for robust feature generation in audiovisual emotion recognition,” ICASSP ’13
  4. A. Krizhevsky, et al., “Imagenet classification with deep convolutional neural networks,” NIPS ’12
  5. A. Y. Hannun, et al., “Deep speech: Scaling up end-to-end speech recognition,” CoRR, vol. abs/1412.5567, 2014.
  6. The Economist. Microphones As Sensors: Teaching Old Microphones New Tricks.

http://www.economist.com/news/technology-quarterly/21578518-sensor-technology-micro

phones-are-designed-capture-sound-they-turn-out.

  1. O. Amft, M. Stäger, P. Lukowicz, G. Tröster. Analysis of Chewing Sounds for Dietary Monitoring. UbiComp ’05, pages 56–72, 2005.
  2. E. C. Larson, T. Lee, S. Liu, M. Rosenfeld, S. N. Patel. Accurate and Privacy Preserving Cough Sensing using A Low-cost Microphone. UbiComp ’11, pages 375–384, 2011.
  3. P. Mohan, V. N. Padmanabhan, R. Ramjee. Nericell: Rich Monitoring of Road and Traffic Conditions using Mobile Smartphones. SenSys ’08, pages 323–336, 2008.
  4. H. Lu, D. Frauendorfer, M. Rabbi, M. S. Mast, G. T. Chittaranjan, A. T. Campbell, D. Gatica-Perez, T. Choudhury. Stresssense: Detecting Stress in Unconstrained Acoustic Environments using Smartphones. UbiComp ’12, pages 351–360, 2012.
  5. K. K. Rachuri, M. Musolesi, C. Mascolo, P. J. Rentfrow, C. Longworth, A. Aucinas. Emotionsense: A Mobile Phones based Adaptive Platform for Experimental Social Psychology Research. Ubicomp ’10, pages 281–290, 2010.
  6. C. Xu, S. Li, G. Liu, Y. Zhang, E. Miluzzo, Y.-F. Chen, J. Li, B. Firner. Crowd++: Unsupervised Speaker Count with Smartphones. UbiComp’13, pages 43–52, 2013
  7. M. Rabbi, S. Ali, T. Choudhury, E. Berke. Passive and In-situ Assessment of Mental and Physical Well-being using Mobile Sensors. Ubicomp ’11, pages 385–394, 2011.
  8. Y. Lee, C. Min, C. Hwang, J. Lee, I. Hwang, Y. Ju, C. Yoo, M. Moon, U. Lee, J. Song. Sociophone: Everyday Face-to-face Interaction Monitoring Platform using Multi-phone Sensor Fusion. MobiSys ’13, pages 375–388, 2013.
  9. Y. Chon, N. D. Lane, Y. Kim, F. Zhao, H. Cha. Understanding the Coverage and Scalability of Place-centric Crowdsensing. UbiComp ’13, pages 3–12, 2013.
  10. L. N. Huynh, R. K. Balan, and Y. Lee, “DeepSense: A GPU-based deep convolutional neural network framework on commodity mobile devices,” in Proc. Workshop Wearable Syst. Appl., 2016, pp. 25–30.
  11. J. E. Stone, D. Gohara, and G. Shi. Opencl: A parallel programming standard for heterogeneous computing systems. Computing in science & engineering, 12(1-3):66–73, 2010.
  12. T. Y.-H. Chen, L. Ravindranath, S. Deng, P. Bahl, and H. Bala krishnan, “Glimpse: Continuous, real-time object recognition on mobile devices,” in Proc. 13th ACM Conf. Embedded Netw. Sensor Syst., 2015, pp. 155–168.
  13. S. Han, H. Shen, M. Philipose, S. Agarwal, A. Wolman, and A. Krishnamurthy, “MCDNN: An approximation-based execution framework for deep stream processing under resource constraints,” in Proc. 14th Annu. Int. Conf. Mobile Syst. Appl. Serv., 2016, pp. 123–136.
  14. N. D. Lane, S. Bhattacharya, P. Georgiev, C. Forlivesi, L. Jiao, L. Qendro, and F. Kawsar, “DeepX: A software accelerator for low-power deep learning inference on mobile devices,” in Proc. 15th ACM/IEEE Int. Conf. Inf. Process. Sensor Netw., 2016, pp. 1–12.
  15. L. N. Huynh, Y. Lee, and R. K. Balan, “DeepMon: Mobile GPU-based deep learning framework for continuous vision applications,” in Proc. 15th Annu. Int. Conf. Mobile Syst. Appl. Serv., 2017, pp. 82–95.
  16. G. E. Hinton, et al., “A fast learning algorithm for deep belief nets,” Neural Comput., vol. 18, no. 7, pp. 1527–1554, Jul. 2006.
  17. Y. LeCun, et al., “Gradient-based learning applied to document recognition,” Proc. of IEEE, vol. 86, no. 11, pp. 2278–2324.
  18. N. Lane, et al., “Can Deep Learning Revolutionize Mobile Sensing?,” HotMobile 2015.
  19. “LG G Watch R,” https://www.qualcomm.com/products/snapdragon/wearables/lg-g-watch-r.
  20. “QualcommSnapdragon400,” https://www.qualcomm.com/products/snapdragon/processors/400.
  21. N. D. Lane, P. Georgiev, and L. Qendro, “DeepEar: Robust smart-phone audio sensing in unconstrained acoustic environments using deep learning,” in Proc. ACM Int. Joint Conf. Pervasive Ubiquitous Comput., 2015, pp. 283–294.
  22. K. K. Rachuri, M. Musolesi, C. Mascolo, P. J. Rentfrow, C. Longworth, A. Aucinas. Emotionsense: A Mobile Phones based Adaptive Platform for Experimental Social Psychology Research. Ubicomp ’10, pages 281–290, 2010.
  23. H. Lu, D. Frauendorfer, M. Rabbi, M. S. Mast, G. T. Chittaranjan, A. T. Campbell, D. Gatica-Perez, T. Choudhury. Stresssense: Detecting Stress in Unconstrained Acoustic Envionments using Smartphones. UbiComp ’12, pages 351–360, 2012.
  24. H. Lu, A. J. B. Brush, B. Priyantha, A. K. Karlson, J. Liu. Speakersense: Energy Efficient Unobtrusive Speaker Identification on Mobile Phones. Pervasive’11, pages 188–205, 2011.
  25. D. A. Reynolds, R. A. Rose. Robust Text-independent Speaker Identification using Gaussian Mixture Speaker Models. Transactions on Speech and Audio Processing, pages 72–83, (3):1, 1995.
  26. H. Lu, J. Yang, Z. Liu, N. Lane, T. Choudhury, A. Campbell. The Jigsaw Continuous Sensing Engine for Mobile Phone Applications. SenSys ’10, pages 71–84, 2010.
  27. J. Ba and R. Caruana, “Do deep nets really need to be deep?” in Proc. 27th Int. Conf. Neural Inf. Process. Syst., 2014, pp. 2654–2662.
  28. N. D. Lane, S. Bhattacharya, P. Georgiev, C. Forlivesi, L. Jiao, L. Qendro, and F. Kawsar, “DeepX: A software accelerator for low-power deep learning inference on mobile devices,” in Proc. 15th ACM/IEEE Int. Conf. Inf. Process. Sensor Netw., 2016, pp. 1–12.
  29. A. Collet Romea, M. Martinez Torres, and S. Srinivasa. The moped framework: Object recognition and pose estimation for manipulation. In IJRR, 2011.
  30. H. Bay, A. Ess, T. Tuytelaars, and L. Van Gool. Speeded-up robust features (surf). In CVIU, 2008.
  31. R.-E. Fan, K.-W. Chang, C.-J. Hsieh, X.-R. Wang, and C.-J. Lin. LIBLINEAR: A Library for Large Linear Classification. In JMLR, 2008.
  32. A. Romero, N. Ballas, S. E. Kahou, A. Chassang, C. Gatta, and Y. Bengio. Fitnets: Hints for thin deep nets. In Proceedings of the International Conference on Learning Representations (ICLR), 2015.
  33. C. Szegedy, W. Liu, Y. Jia, P. Sermanet, S. Reed, D. Anguelov, D. Erhan, V. Vanhoucke, and A. Rabinovich. Going deeper with convolutions. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2015.
  34. S. Han, H. Mao, and W. J. Dally. Deep compression: Compressing deep neural network with pruning, trained quantization and huffman coding. In Proceedings of the International Conference on Learning Representations (ICLR), 2016.
  1. M. Rastegari, V. Ordonez, J. Redmon, and A. Farhadi. XNOR-Net: ImageNet Classification Using Binary Convolutional Neural Networks. arXiv:1603.05279, Mar. 2016.
  2. M. Rastegari, V. Ordonez, J. Redmon, and A. Farhadi. XNOR Net: ImageNet Classification Using Binary Convolutional Neural Networks. arXiv:1603.05279, Mar. 2016.
  3. Google glass. https://developers.google.com/glass/. Accessed: 2016-12-8.
  4. Hololens. https://www.microsoft.com/microsoft-hololens/en-us.
  5. Chen, T. Y.-H., Ravindranath, L., Deng, S., Bahl, P., and Balakrishnan, H. Glimpse: Continuous, real-time object recognition on mobile devices. Proceedings of the 13th ACM Conference on Embedded Networked Sensor Systems, SenSys ’15, pages 155–168, New York, NY, USA, 2015. ACM.
  1. Boger, J., Hoey, J., Poupart, P., Boutilier, C., Fernie, G., and Mihailidis, A. A planning system based on markov decision processes to guide people with dementia through activities of daily living. IEEE Transactions on Information Technology in Biomedicine, 10(2):323–333, 2006.
  2. Chao wu , qiushi li, ziyan fu, wenwu zhu , yaoxue zhang , enabling flexible resource allocation in mobile deep learning systems, IEEE transactions on parallel and distributed systems, vol. 30, no. 2, february 2019.