You are on page 1of 4

1

Review of Machine Learning And Its


Hardware Platform
Xiaotian Liu

Abstract—As the gordian technique of Artificial Intelligence, Machine Learning is,by far,the most effective way to make
computer acquire intelligence. In addition to the algorithm, the key to the development of machine learning is to create
corresponding hardwares. And we hope the hardwares should be suitable for the operation mode of Machine Learning
and can be integrated into the algorithm system of Machine Learning. So in this paper, we will talk about how to support
the progress of software and solve several problems about Machine Learning through the development of hardware,
which shows the importance of hardware platform in Machine Learning.

Keywords: Machine Learning; Hardware Platform; GPU; FGPA

1 I NTRODUCTION creating hardwares that can support local Ma-


chine Learning computing becomes a break-
W ith the development in these years, the
key technologies of Artificial Intelligence
have changed greatly from classical Artificial
through point in the development of Machine
Learning.
So in this paper,firstly,we will briefly intro-
Intelligence to Artificial Neural Network. How- duce the basic concepts and process of Machine
ever, the core idea of Artificial Intelligence al- Learning, then we will introduce several dif-
ways keeps the same ,which is Machine Learn- ficulties encountered by Machine Learning in
ing.Because of its rapid development in recent hardwares and talk about several corresspod-
years, Machine Learning has gradually formed ing solutions.
a series of mature and feasible algorithms at
the software level, such as Decision Tree, Clus-
tering, Bayesian Classification, Support Vector
Machine, EM, Adaboost and so on.At the hard-
ware level, current machine learning is mainly 2 BASIC CONCEPTS OF M ACHINE
realized by ASIC1 , FPGA2 and GPU3 .
In the era of big data, with such a high speed L EARNING
of data production and generation, method of
sending original data to the cloud for com- Machine learning is mainly divided into
puting is gradually outdated, and there is an four steps, which are obtaining data sources,
increasing demand for local analysis and calcu- data analysis ,feature selection, training and
lation near sensors.Embedding Machine Learn- evaluation of data sets. Details are shown in
ing on hardware platforms, also solves some the Table 1:
privacy and potential security issues.Therefore,
Table 1:Four steps of machine learning
• Xiaotian Liu is with the Department of Computer Science Steps Details
and Electronic Engineering, Peking University, Beijing, China,
1800012723. The quality and quantity of data collected will
directly determine whether the prediction mo-
del can be built.Besides,the sparsity and attrib-
Data source
1. ASIC:Application Specific Integrated Circuit utes of the data from the data source can also
influence the design of the algorithms of Mach-
2. FPGA:Field-Programmable Gate Array ine Learning.
3. GPU:Graphics Processing Unit
2

Table 1:Four steps of machine learning


are using GPUs for deep learning training, and
Steps Details NVIDIA dominates the current deep learning
Data analysis mainly focuses on data disco- market with its massively parallel GPUs and
very and visualization. The main content is
Data analysis
to find the relationship between independe- CUDA, a dedicated GPU programming archi-
nt variables in the- tecture.
data and determine the correlation coefficie-
nt. In addition to GPUs, ASIC and FPGA also
The main purpose of feature selection and provide some other solutions for Deep Learn-
feature extraction is to transform the origi- ing hardware implementation. In addition,
Feature selection
nal data into meaningful input for a given
task. more and more companies have developed
By using the well processed data set to trai- specialized hardwares for Deep Learning, such
n the established model, the training proce- as Google’s TPU/Tensor Processing Unit and
ss is actually the process of parameter mod-
Data sets training
ulation.The model with high fitting degree Intel’s Xeon Phi Knight’s Landing.
can be obtained by selecting the best algor-
ithm of training.
3.2 Problems faced by machine learning at
the hardware level
Today’s more advanced Machine Learning
method,called deep neural networks (DNN), As we have mentioned earlier, Machine
does not need to use manually annotated fea- Learning is becoming more and more embed-
tures It gets the features from the data auto- ded.Key metrics for embedded Machine Learn-
matically .This automatic learning method can ing are accuracy, power consumption, through-
map input data to high-dimensional space, so put/latency, and cost.
as to obtain more accurate calculation results, For Machine Learning hardware systems,
at the cost of higher complexity. programmability is the first thing to be consid-
ered, because when the conditions, such as the
problem, data environment and so on, change,
3 H ARDWARE PLATFORMS OF MA - the weight coefficient of the corresponding
CHINE LEARNING nodes should also change.And the requirement
of programmability will result in the increase
3.1 Current usage of machine learning
of data calculation and data transfer.Therefore,
hardware
how to increase the computing power of Ma-
Take deep learning as an example. Baidu’s chine Learning hardware and minimize data
Silicon Valley Artificial Intelligence Laboratory transmission becomes the key to develop ma-
(SVAIL) has proposed a DeepBench bench- chine learning hardware.
mark for deep learning hardware, which fo- In addition, throughput depends on the com-
cuses on measuring the hardware performance putation. How to process and analyze data to
of basic calculations rather than the perfor- reduce the amount of computation is also a key
mance of learning models.Based on Convo- problem of developing hardware systems.
lution Neural Network (CNN) and Recur- In addition, from the perspective of industry,
rent Neural Network(RNN),DeepBench puts how to make machine learning hardwares gain
forward four basic operations: Matrix Multi- industrial value( in other words,how to reduce
plication (Matrix Multiplication), Convolution its cost and power consumption) is also a prob-
(Convolution), Circulation Layer (Recurrent the lem needed to be solved in the development of
Layers), and All Reduce4 . machine learning hardware.
In terms of the computing methods we
have mentioned above, GPU has found a new 3.3 Several solutions to the hardware prob-
growth point in Machine Learning due to its lems in machine learning
ability to complete computation independently
and execute parallel commands.In fact, most 3.3.1 Parallel framework of machine learning
Deep Learning (Machine Learning) researchers As a pioneering creation in the field of Ma-
chine Learning, the Convolutional Neural Net-
4. All Reduce: Global Reduction Operations work modeled after the visual cortex can be
3

used in both supervised learning and unsuper- 3.3.2 Application of new nonvolatile storage
vised learning.Because of its deep structure,ut’s based on memristor in machine learning
suitable for tasks with natural hierarchical fea- Since the process of Machine Learning needs
tures, such as voice, images, text, etc. In addi- large amounts of data, the ability to store,
tion, the representability of convolutional neu- read and write data has a significant impact
ral network makes it superior in completing the on the effectiveness of Machine Learning.The
iteration work. new non-volatile storage based on memristor
Aiming at this deep model, hardware espe- takes advantage of the fact that the resistance of
cially designed for this unique computing net- memristor can ”remember” the information of
work has gradually emerged.For example, the the charge of memristor to gain the character-
CNN FPGA accelerator with high throughput, istic of non-volatile.Due to this special feature
which has been developed in recent years, is of memristor, the non-volatile storage based on
designed to achieve the highest efficiency in the memristor has the characteristics of fast read-
calculation of Convolutional Neural Network. ing speed and small storage units.These char-
The main principle is shown in Figure 1. acteristics give this new non-volatile memory
unique advantages in Machine Learning.
In addition, some researches focus on the
technology of integrating multiplication di-
rectly on advanced non-volatile storage, in
which conductance is taken as the weight coef-
ficient of every node in Deep Learning, voltage
is taken as input and current is taken as out-
put.Therefore, such an ohm’s law process com-
pletes a multiplication operation.This method
can greatly improve the efficiency of multipli-
cation calculation in Machine Learning after
proper weighting processing.

4 S UMMARY OF MACHINE LEARNING


HARDWARE PLATFORMS
Machine learning, as an important realiz-
ing method of Artificial Intelligence, needs not
only algorithm improvement and optimization,
but also hardware platform support and in-
novation.For Machine Learning, considerations
Fig. 1: Top-Level Architecture of the
on hardwarse mainly focus on the throughput,
Convolutional Neural Network Accelerator
accuracy, latency, power consumption and the
cost.In view of these problems, we respectively
put forward two effective solutions, which has
We can see from the picture that the pro- been proved to be useful and promising. In
cessing unit (PE) was divided into different a word, the innovation and development of
Arraies, which can attribute data from DRAM Machine Learning hardwares are inseparable
to differernt PE Arraies. And these indepen- from softwares.The creation of machine learn-
dent computing units can be efficient in the ing hardware should become increasingly spe-
dot product calculation.After calculation, the cialized, adapting to different algorithms or
results are sent to a specific position waiting tasks to achieve higher efficiency.While requir-
for next layer’s calculation.This process shows ing performance, Machine Learning hardware
exactly how the design of hardware improves manufacturing should also take full account of
the computational efficiency. power consumption, cost and other industrial
4

issues.More researches and efforts are needed


to solve these problems of machine learning
hardware.

ACKNOWLEDGMENTS
The author would like to thank the blue sky
and the white cloud and himself.

R EFERENCES
[1] Hardware for Machine Learning:Challenges and
Opportunities,Vivienne Sze,Yu-Hsin,Joel Emer,Amr
Suleiman,Zhengdong Zhang,MIT,Cambridge,MA 02139
[2] Recent advances in convolutional neural net-
works,Gu,J.Wang, Z.Kuen, J.Ma, L.Shahroudy, A.Shuai,
B.Liu, T.Wang, X.Wang, L.Wang, 2015,arXiv preprint
arXiv:1512.07108.
[3] Nanoscale Memristor Device as Synapse in Neuromorphic
Systems,Nano Letters,2010.03.01.

You might also like