没有合适的资源?快使用搜索试试~ 我知道了~
首页GPU驱动的深度学习:图像与特征识别关键技术
"本研究论文《深度学习与GPU技术在图像与特征识别中的应用》由Alison B. Lowndes撰写,旨在探讨和实证使用图形处理器(GPU)驱动的深度学习神经网络在图像特征识别领域的潜力。论文针对人工智能领域的学生和研究人员,提供如何运行卷积神经网络(CNN)以及理解该领域其他技术的指导。 论文的核心内容聚焦于利用GPU的优势,因为它们能够在处理大量数据和执行复杂计算时展现高效性能。通过GPU加速,深度学习算法,尤其是CNN,能够快速地处理图像数据,执行卷积、池化等操作,从而实现高精度的特征检测和识别。这种技术的进步使得认知计算系统,如IBM的Watson,能够通过概率自然语言处理与用户进行更自然的交互,推动了人工智能领域的快速发展。 NVIDIA甚至推出了一个即插即用的CNN“设备”,整合了4个GPU和其DIGITS工具包,这进一步简化了深度学习实践者的工作流程,使得普通人也能轻易上手深度学习项目。这项研究不仅展示了GPU技术在深度学习中的关键作用,也强调了大规模数据集和优化算法在驱动深度学习复兴中的不可替代性。 此篇论文不仅提供了深度学习在图像特征识别中的具体实现方法,还揭示了GPU技术如何成为这一领域不可或缺的技术基础设施,对于那些希望在这个快速发展的领域中探索或应用深度学习的人来说,具有很高的参考价值。作者期待通过这篇研究,促使Leeds大学等机构在未来加大对GPU技术的投资,以推动人工智能领域的持续创新和发展。"
资源详情
资源推荐
![](https://csdnimg.cn/release/download_crawler_static/9896567/bg10.jpg)
This is an important step toward making computers much more brainlike than ever
before, but perhaps trying to mimic the brain, will not be the best route forward [
24
]
much the same way that aviation engineers realised it was much better to utilise
concepts like lift than to try to mimic feathers for aircraft. Professor Michael Jordan
states "There is nothing very neural about neural networks ... no spikes in deep-learning
systems.. no dendrites” [
24
]. Regarding backpropagation Jordan describes how the brain
does not operate the same way, though he acknowledges it led to significant progress.
Numenta’s Jeff Hawkins argues that “intelligent machines will be based on models of the
5
neocortex, not the rest of the brain”, there is a big difference. The most recent paper to
have been accepted into the Computer Vision and Pattern Recognition conference,
CVPR2015, describes how differently CNNs see the world than humans and how this can
also lead to inaccuracy. Nguyen, Clune et al of the University of Wyoming show that
CNNs, if trained with “fooled” images will give a majority false positive results. The
technique is demonstrated on www.picbreeder.org where the user is the fitness function.
The paper illustrates the significance that CNNs can be trained on whatever we want
them to be trained on, and that we can even flip the linear processes; fixed image,
tweaked parameters v. fixed parameters, altered image. The flaw itself, is inherited from
the linear nature of classifiers themselves. The CNN is trained via a repeated process of
sampling data, calculating gradients and updating. One gradient updates one parameter
to increase the correct image score, along with all other gradient-parameter updates.
This is repeated epoch after epoch on all images in the training data until a suitable
accuracy is obtained. Perhaps more efficient methods lie within self-organisation and in
the random way the brain seems to construct its own neural network. Whatever the
solution, today, the most successful artificial neural networks are deep and
multi-layered.
1.4 Neuromorphic chips
Neuromorphic chips are currently being funded by DARPA via HRL Laboratories and IBM,
to enhance AI and avoid Moores Law and its fundamental performance limit. Deep
learning will soon integrate with mobile phones due to Qualcomm bundling its Zeroth
software with its Snapdragon 820 processor - to recognize speech, music, scenes, pick
out objects and even spot patterns of behaviour via the device’s sensors.
5
http://recode.net/2015/03/02/theterminatorisnotcomingthefuturewillthankus/
16
![](https://csdnimg.cn/release/download_crawler_static/9896567/bg11.jpg)
Fig 2. Neuromorphic chip designed by Heidelberg University. The chip features 384 neurons, 100,000 synapses
and operates at a speed of approx. 100,000x biological real-time (Human Brain Project, May 2013)
The Human Brain Project in Europe has spent EU100 Million on neuromorphics and
many laboratories are attempting to reconstruct a silicon brain. Neural processing units
(NPU) such as IBM’s SyNapse chip, unveiled in August 2014, are designed to be extremely
energy-efficient and optimized for sensory input processing. Testing showed the chips
perform 100x faster, consuming 100,000x less power than a laptop running standard von
Neumann architecture and CPU's [
27
]. NPUs do not separate memory and processing
blocks as in von Neumann architecture so bottlenecks and latency are vastly reduced
with neurons and synapses intertwining both functions. Unfortunately these chips also
require a completely new approach to programming, being developed by Rice University
in the PLINY Project [
20
].
2.0Objectives
This report provides a general introduction to the topic of deep learning with a history of
the field leading to current progress. A short introduction to image recognition follows,
prior to fuller definition of the different flavours of neural networks and techniques used
in deep learning, with a number of specific applications in both academia and industry.
From section 5 the role of GPU's are introduced, in depth, with specific testing further
explained. Various frameworks, environments and software are tested before leading
17
![](https://csdnimg.cn/release/download_crawler_static/9896567/bg12.jpg)
into the author's key interest of the use of deep learning in healthcare, specifically
recognition of subtle features in histology. There is attention given to the mathematics
behind the success of the current state-of-the-art algorithms in use before a conclusion
with details of future work. A substantial list of references and full literature review is
provided with multiple appendices should the reader wish to duplicate any tests made or
simply to dive into the fascinating field of deep learning with convolutional neural
networks. Appendices include actual quotations and advice for obtaining the hardware
necessary though it should be noted that this report leaves with it a 2880 core, 12GB GPU
currently housed at
csfyp01
in the DEC-10 machine room, thanks to the NVIDIA
Corporation.
The report will cover :
➢ an overview of the current state-of-the-art deep learning techniques available for
natural image recognition and mitosis (feature) detection in medical images.
➢ research and practical implementation of GPGPU (general purpose
programmability on a GPU).
➢ research and practical implementation of neural networks with comparison of
various machine learning techniques.
➢ comparison between CPU and GPU capability and performance.
➢ application of deep learning for recognition of features and objects in natural and
medical images.
➢ presentation of use-case to persuade the School of Computing to invest in GPUs
and assist current and future work in many fields of research.
2.1 Personal Statement
[
removed for confidentiality
]
Timing and numerical results detailed at section 11 with extensive details of testing in
the appendices.
18
![](https://csdnimg.cn/release/download_crawler_static/9896567/bg13.jpg)
3.0Imagerecognition
It is important to understand the semantic gap between what "we see" and what a
computer reads. An image of a cute cat implies knowledge, recognition, feelings (‘cute’). A
computer sees a patch of integers between [0, 255], in 3-dimensions representing the
Red, Green and Blue colour channels. All neural net activations are arranged in 3D
(height x width x depth). Images from the CIFAR-10 dataset are represented as 32 width x
32 height x 3 depth (RGB) per image with 10 classes [
airplane, automobile, bird, cat, deer,
dog, frog, horse, ship, truck
]. The central pixel of a square patch of image data tends to
be used to determine the class of the image and the network is trained on these patches
(of known class). The 10 classes each label 50,000 training images used to train a DL
system. Once trained the system is tested with a further 10,000 unlabelled test images.
Challenges that humans take for granted, a computer needs help with; pose, viewpoint,
lighting, scale, deformation, occlusion, noise, clutter, intra-class variation e.g 100
different types of chair. All these are big problems that neural nets have largely solved
since the data-driven approach.
From the mathematical representation algorithms simply calculate f(x
i
, Weight, bias) as a
matrix Wx
i
and a vector b. The trained network can then classify new, unseen, images
and trivially detect objects within by comparing pixels close to the object centroid with
the different background pixels [
52
].
Fig 3. Conversion to digital Li et al, Stanford
Nearest neighbour classifiers can predict the probability of an unlabeled image based on
the knowledge gained from the training set in the form of hyperparameters like distance,
19
![](https://csdnimg.cn/release/download_crawler_static/9896567/bg14.jpg)
pixel-wise abs() or Euclidean for example. See Appendix J for full code breakdown of
image preprocessing on the Street View House Numbers dataset (Google).
4.0NeuralNets
The area of deep learning spans a wide spectrum (covered):
MODEL
DEEP
SHALLOW
Unsupervised
Sparse Autoencoder
LSTM
Deep Belief Network
Bayesian NP
Restricted Boltzmann
Machines
Supervised
Neural Networks (MLP)
Convolutional NNs
Recurrent NNs
Perceptron
SVM
4.1 Perceptrons
Perceptrons are a close artificial implementation of the basic computational unit of the
brain; the biological neuron. A biological neuron differs in one significant way to its
artificial representation. Unlike the perceptron, a biological neuron is capable of
recognising 100's of different patterns, not just one added to others, summed by weights,
modified by weights, etc.
Perceptrons can, however, concatenate into more complex neural networks. The synaptic
strength (weights) in a perceptron are learnable and control the strength (excitatory =
positive weight or inhibitory = negative weight) of one neuron on another. In the brain,
dendrites carry signals to the soma (cell body) where they are summed. If the final sum
is above a certain threshold, the neuron can fire, sending an electrical ‘spike’ along its
axon. In the perceptron, only the frequency of the firing (rate code) communicates
information, modelled by an activation function [
11
]. This differs to how the actual
neocortex works, which is out of scope of this report. Nevertheless it is an important
point, explained further in the context of Numenta’s work in Appendix N.
Python implementation:
class Neuron:
20
剩余136页未读,继续阅读
![pdf](https://img-home.csdnimg.cn/images/20210720083512.png)
![pdf](https://img-home.csdnimg.cn/images/20210720083512.png)
![](https://csdnimg.cn/download_wenku/file_type_ask_c1.png)
![](https://csdnimg.cn/download_wenku/file_type_ask_c1.png)
![](https://csdnimg.cn/download_wenku/file_type_ask_c1.png)
![](https://csdnimg.cn/download_wenku/file_type_ask_c1.png)
![](https://csdnimg.cn/download_wenku/file_type_ask_c1.png)
![](https://csdnimg.cn/download_wenku/file_type_ask_c1.png)
![](https://csdnimg.cn/download_wenku/file_type_ask_c1.png)
![](https://csdnimg.cn/download_wenku/file_type_ask_c1.png)
![](https://csdnimg.cn/download_wenku/file_type_ask_c1.png)
![](https://csdnimg.cn/download_wenku/file_type_ask_c1.png)
![](https://csdnimg.cn/download_wenku/file_type_ask_c1.png)
![](https://csdnimg.cn/download_wenku/file_type_ask_c1.png)
![](https://csdnimg.cn/download_wenku/file_type_ask_c1.png)
![](https://csdnimg.cn/download_wenku/file_type_ask_c1.png)
![](https://csdnimg.cn/download_wenku/file_type_ask_c1.png)
![](https://csdnimg.cn/download_wenku/file_type_ask_c1.png)
![](https://csdnimg.cn/download_wenku/file_type_ask_c1.png)
![](https://profile-avatar.csdnimg.cn/default.jpg!1)
hqhsc
- 粉丝: 0
- 资源: 3
上传资源 快速赚钱
我的内容管理 展开
我的资源 快来上传第一个资源
我的收益
登录查看自己的收益我的积分 登录查看自己的积分
我的C币 登录后查看C币余额
我的收藏
我的下载
下载帮助
![](https://csdnimg.cn/release/wenkucmsfe/public/img/voice.245cc511.png)
最新资源
- Lombok 快速入门与注解详解
- SpringSecurity实战:声明式安全控制框架解析
- XML基础教程:从数据传输到存储解析
- Matlab实现图像空间平移与镜像变换示例
- Python流程控制与运算符详解
- Python基础:类型转换与循环语句
- 辰科CD-6024-4控制器说明书:LED亮度调节与触发功能解析
- AE particular插件全面解析:英汉对照与关键参数
- Shell脚本实践:创建tar包、字符串累加与简易运算器
- TMS320F28335:浮点处理器与ADC详解
- 互联网基础与结构解析:从ARPANET到多层次ISP
- Redhat系统中构建与Windows共享的Samba服务器实战
- microPython编程指南:从入门到实践
- 数据结构实验:顺序构建并遍历链表
- NVIDIA TX2系统安装与恢复指南
- C语言实现贪吃蛇游戏基础代码
资源上传下载、课程学习等过程中有任何疑问或建议,欢迎提出宝贵意见哦~我们会及时处理!
点击此处反馈
![](https://img-home.csdnimg.cn/images/20220527035711.png)
![](https://img-home.csdnimg.cn/images/20220527035711.png)
![](https://img-home.csdnimg.cn/images/20220527035111.png)
安全验证
文档复制为VIP权益,开通VIP直接复制
![](https://csdnimg.cn/release/wenkucmsfe/public/img/green-success.6a4acb44.png)