deep learning case study pdf

Understanding Deep Learning: Case Study Based Approach

• First Online: 28 January 2021

Cite this chapter

• Manisha Galphade 4 ,
• Nilkamal More 4 ,
• V. B. Nikam 4 ,
• Biplab Banerjee 5 &
• Arvind W. Kiwelekar 6  

Part of the book series: EAI/Springer Innovations in Communication and Computing ((EAISICC))

743 Accesses

Deep learning is a much focused domain of artificial neural networks. Deep learning algorithms try to learn massive amounts of unlabelled data and make a better analysis. With deep learning, all layers learn the input data and transform it into a more abstract and composite format. The word “deep” means higher numbers of hidden layers in which the data from one layer to another is transformed to generate the most accurate outcome. Deep learning architecture has been applied to different fields like medical image analysis, machine translation, bioinformatics, speech recognition, social network filtering, computer vision, audio recognition drug design, natural language processing, and so on. This chapter discusses important deep learning applications across different disciplines, their contribution to the real world, and a study of the architectures and methods used by each application. This chapter also introduces the differences between machine learning and deep learning. Finally, this chapter concludes with future aspects and conclusions.

This is a preview of subscription content, log in via an institution to check access.

Access this chapter

Subscribe and save.

• Get 10 units per month
• Download Article/Chapter or eBook
• 1 Unit = 1 Article or 1 Chapter
• Cancel anytime
• Available as PDF
• Read on any device
• Instant download
• Own it forever
• Available as EPUB and PDF
• Compact, lightweight edition
• Dispatched in 3 to 5 business days
• Free shipping worldwide - see info
• Durable hardcover edition

Tax calculation will be finalised at checkout

Purchases are for personal use only

Institutional subscriptions

Similar content being viewed by others

Deep Learning Architectures, Methods, and Frameworks: A Review

Introduction to Deep Learning

Deep Learning Techniques: An Overview

W.G. Hatcher, W. Yu, A survey of deep learning: Platforms, applications and emerging research trends. IEEE Access 6 , 24411–24432 (2018). https://doi.org/10.1109/access.2018.2830661

Article   Google Scholar  

G. Litjens, T. Kooi, B.E. Bejnordi, A.A.A. Setio, F. Ciompi, M. Ghafoorian, C.I. Sánchez, A survey on deep learning in medical image analysis. Med. Image Anal. 42 , 60–88 (2017). https://doi.org/10.1016/j.media.2017.07.005

T. Brosch, R. Tam, Manifold learning of brain MRIs by deep learning. Lect. Notes Comput. Sci, 633–640 (2013). https://doi.org/10.1007/978-3-642-40763-5_78

A. Teramoto, H. Fujita, O. Yamamuro, T. Tamaki, Automated detection of pulmonary nodules in PET/CT images: Ensemble false-positive reduction using a convolutional neural network technique. Med. Phys. 43 (6Part1), 2821–2827 (2016). https://doi.org/10.1118/1.4948498

Q. Dou, H. Chen, L. Yu, L. Zhao, J. Qin, D. Wang, P.-A. Heng, Automatic detection of cerebral microbleeds from MR images via 3D convolutional neural networks. IEEE Trans. Med. Imaging 35 (5), 1182–1195 (2016). https://doi.org/10.1109/tmi.2016.2528129

V. Jindal (2016). Integrating mobile and cloud for PPG signal selection to monitor heart rate during intensive physical exercise. Proceedings of the International Workshop on Mobile Software Engineering and Systems—MOBILESoft ‘16. doi: https://doi.org/10.1145/2897073.2897132

D.R. Kelley, J. Snoek, J.L. Rinn, Basset: Learning the regulatory code of the accessible genome with deep convolutional neural networks. Genome Res. 26 (7), 990–999 (2016). https://doi.org/10.1101/gr.200535.115

Krizhevsky, I. Sutskever, and G. Hinton, Imagenet classification with deep convolutional neural networks, In Proceedings of Neural Information Processing Systems (NIPS) (2012)

Google Scholar  

E.-J. Ong, S. Husain, M. Bober-Irizar, M. Bober, Deep architectures and ensembles for semantic video classification. IEEE Transactions on Circuits and Systems for Video Technology 29 , 1–1 (2018). https://doi.org/10.1109/tcsvt.2018.2881842

K.G. Srinivasa, S. Anupindi, R. Sharath, & S.K. Chaitanya, Analysis of facial expressiveness captured in reaction to videos. 2017 IEEE 7th International Advance Computing Conference (IACC) (2017) https://doi.org/10.1109/iacc.2017.0140

T.N. Sainath, R.J. Weiss, K.W. Wilson, B. Li, A. Narayanan, E. Variani, et al., Multichannel signal processing with deep neural networks for automatic speech recognition. IEEE/ACM Trans. Audio Speech Lang. Process. 25 (5), 965–979 (2017). https://doi.org/10.1109/taslp.2017.2672401

Y. N. Dauphin, A. Fan, M. Auli, and D. Grangier, Language modeling with gated convolutional networks, in arXiv Preprint, arXiv:1612.08083 (2016)

O. Araque, I. Corcuera-Platas, J.F. Sánchez-Rada, C.A. Iglesias, Enhancing deep learning sentiment analysis with ensemble techniques in social applications. Expert Syst. Appl. 77 , 236–246 (2017). https://doi.org/10.1016/j.eswa.2017.02.002

A. Severyn, & A. Moschitti, Learning to Rank Short Text Pairs with Convolutional Deep Neural Networks. Proceedings of the 38th International ACM SIGIR Conference on Research and Development in Information Retrieval - SIGIR ‘15 (2015) doi: https://doi.org/10.1145/2766462.2767738

S. Liu, N. Yang, M. Li, and M. Zhou, A recursive recurrent neural network for statistical machine translation, in Proceeding 52nd Annual Meeting Association Computational Linguistics, pp.1491–1500 (2014)

H. Schwenk, Continuous-space language models for statistical machine translation. Prague Bull. Math. Ling. 93 (1) (2010). https://doi.org/10.2478/v10108-010-0014-6

E. Busseti, I. Osband, and S. Wong, Deep learning for time series modelling, Technical report, Stanford University (2012)

B. Batres-Estrada, Deep learning for multivariate financial time series (Dissertation) (2015). http://urn.kb.se/resolve?urn=urn:nbn:se:kth:diva-168751

X. Dong, L. Qian, Y. Guan, L. Huang, Q. Yu and J. Yang, A multiclass classification method based on deep learning for named entity recognition in electronic medical records, 2016 New York Scientific Data Summit (NYSDS), New York, pp. 1–10 (2016). doi: https://doi.org/10.1109/NYSDS.2016.7747810

Download references

Acknowledgments

The authors take this opportunity to thank Faculty Development Centre (VJTI-DBATU) in Geoinformatics, Spatial Computing and Big Data Analytics developed under agencies PMMMNMTT, MHRD, Government of India, New Delhi for capacity building in the domain of Geoinformatics, AI & Machine Learning, Big Data Analytics, Deep Learning, and related domains, which helped us to take up this study ahead.

Author information

Authors and affiliations.

Department of Computer Engineering & IT, Veermata Jijabai Technological Institute, Mumbai, India

Manisha Galphade, Nilkamal More & V. B. Nikam

Center of Studies in Resources Engineering, IIT Bombay, Mumbai, India

Biplab Banerjee

Department of Information Technology, Dr Babasaheb Ambedkar Technological University, Lonere, Raigad, India

Arvind W. Kiwelekar

You can also search for this author in PubMed   Google Scholar

Editor information

Editors and affiliations.

Department of Computer Science and Engineering, SRM Institute of Science and Technology, Kattankulathur, Chennai, Tamil Nadu, India

Polytechnic Institute of Viana do Castel, Viana do Castelo, Portugal

Rights and permissions

Reprints and permissions

Copyright information

© 2021 Springer Nature Switzerland AG

About this chapter

Galphade, M., More, N., Nikam, V.B., Banerjee, B., Kiwelekar, A.W. (2021). Understanding Deep Learning: Case Study Based Approach. In: Suresh, A., Paiva, S. (eds) Deep Learning and Edge Computing Solutions for High Performance Computing. EAI/Springer Innovations in Communication and Computing. Springer, Cham. https://doi.org/10.1007/978-3-030-60265-9_9

Download citation

DOI : https://doi.org/10.1007/978-3-030-60265-9_9

Published : 28 January 2021

Publisher Name : Springer, Cham

Print ISBN : 978-3-030-60264-2

Online ISBN : 978-3-030-60265-9

eBook Packages : Engineering Engineering (R0)

Share this chapter

Anyone you share the following link with will be able to read this content:

Sorry, a shareable link is not currently available for this article.

Provided by the Springer Nature SharedIt content-sharing initiative

• Publish with us

Policies and ethics

• Find a journal
• Track your research

Academia.edu no longer supports Internet Explorer.

To browse Academia.edu and the wider internet faster and more securely, please take a few seconds to  upgrade your browser .

Enter the email address you signed up with and we'll email you a reset link.

• We're Hiring!
• Help Center

Deep Learning Neural Networks: Design and Case Studies

Deep Learning Neural Networks is the fastest growing field in machine learning. It serves as a powerful computational tool for solving prediction, decision, diagnosis, detection and decision problems based on a well-defined computational architecture. It has been successfully applied to a broad field of applications ranging from computer security, speech recognition, image and video recognition to industrial fault detection, medical diagnostics and finance. This comprehensive textbook is the first in the new emerging field. Numerous case studies are succinctly demonstrated in the text. It is intended for use as a one-semester graduate-level university text and as a textbook for research and development establishments in industry, medicine and financial research. Readership: Researchers, academics, professionals, graduate and undergraduate students in machine learning, artificial intelligence, neural networks/networking, software engineering, and in their applications in medicine, se...

Related Papers

Neural Computing and Applications

Lazaros Iliadis

Archives of Computational Methods in Engineering

Munish Jindal

https://www.igi-global.com/

Dr. Priyanka Patel , Priyanka Patel

The chapter is about deep learning fundaments and its recent trends. The chapter mentions many advanced applications and deep learning models and networks to easily solve those applications in a very smart way. Discussion of some techniques for computer vision problem and how to solve with deep learning approach are included. After taking fundamental knowledge of the background theory, one can create or solve applications. The current state-of-the-art of deep learning for education, healthcare, agriculture, industrial, organizations, and research and development applications are very fast growing. The chapter is about types of learning in a deep learning approach, what kind of data set one can be required, and what kind of hardware facility is required for the particular complex problem. For unsupervised learning problems, Deep learning algorithms have been designed, but in the same way Deep learning is also solving the supervised learning problems for a wide variety of tasks.

Saptarshi Sengupta

Deep learning has taken over - both in problems beyond the realm of traditional, hand-crafted machine learning paradigms as well as in capturing the imagination of the practitioner sitting on top of petabytes of data. While the public perception about the efficacy of deep neural architectures in complex pattern recognition tasks grows, sequentially up-to-date primers on the current state of affairs must follow. In this review, we seek to present a refresher of the many different stacked, connectionist networks that make up the deep learning architectures followed by automatic architecture optimization protocols using multiagent approaches. Further, since guaranteeing system uptime is fast becoming an indispensable asset across multiple industrial modalities, we include an investigative section on testing neural networks for fault detection and subsequent mitigation. This is followed by an exploratory survey of several application areas where deep learning has emerged as a game-changing technology - be it anomalous behavior detection in financial applications or financial time-series forecasting, predictive and prescriptive analytics, medical imaging, natural language processing or power systems research. The thrust of this review is on outlining emerging areas of application-oriented research within the deep learning community as well as to provide a handy reference to researchers seeking to embrace deep learning in their work for what it is: statistical pattern recognizers with unparalleled hierarchical structure learning capacity with the ability to scale with information.

Amir Mosavi

Deep learning is an emerging area of machine learning (ML) research. It comprises multiple hidden layers of artificial neural networks. The deep learning methodology applies nonlinear transformations and model abstractions of high level in large databases. The recent advancements in deep learning architec-tures within numerous fields have already provided significant contributions in artificial intelligence. This article presents a state of the art survey on the contributions and the novel applications of deep learning. The following review chronologically presents how and in what major applications deep learning algorithms have been utilized. Furthermore, the superior and beneficial of the deep learning methodology and its hierarchy in layers and nonlinear operations are presented and compared with the more conventional algorithms in the common applications. The state of the art survey further provides a general overview on the novel concept and the ever-increasing advantages and popularity of deep learning.

International Journal for Research in Applied Science and Engineering Technology -IJRASET

IJRASET Publication

Artificial intelligence (AI) is countered to be one of the most trusted techniques to cope with variety of issues. Researchers are delving deeper by using the techniques of AI, such as Machine learning (ML) and deep learning (DL). ML has attained a high attraction in the industry and it is utilized by many applications. Due to drastic increase in data, these techniques are becoming popular amongst the researchers. Long with this, deep learning is the branch of ML which outperformed the conventional techniques of machine learning. This paper presents the brief account on ML and DL. It reviewed how machine and deep learning are utilized and perform different operations. Along with this, a literature survey is presented on the basis of three different domains: Security, Health Management and Big data. This paper gives an overview on Machine learning and deep learning along with the work proposed in this domain.

SN Computer Science, Springer

Dr. Iqbal H. Sarker

Deep learning (DL), a branch of machine learning (ML) and artificial intelligence (AI) is nowadays considered as a core technology of today's Fourth Industrial Revolution (4IR or Industry 4.0). Due to its learning capabilities from data, DL technology originated from artificial neural network (ANN), has become a hot topic in the context of computing, and is widely applied in various application areas like healthcare, visual recognition, text analytics, cybersecurity, and many more. However, building an appropriate DL model is a challenging task, due to the dynamic nature and variations in real-world problems and data. Moreover, the lack of core understanding turns DL methods into black-box machines that hamper development at the standard level. This article presents a structured and comprehensive view on DL techniques including a taxonomy considering various types of real-world tasks like supervised or unsupervised. In our taxonomy, we take into account deep networks for supervised or discriminative learning, unsupervised or generative learning as well as hybrid learning and relevant others. We also summarize real-world application areas where deep learning techniques can be used. Finally, we point out ten potential aspects for future generation DL modeling with research directions. Overall, this article aims to draw a big picture on DL modeling that can be used as a reference guide for both academia and industry professionals.

Asian Journal of Applied Sciences

Maad M . Mijwil

Deep learning has become a favoured trend in many applications serving humanity in the past few years. Since deep learning seeks useful investigation and can learn and train huge amounts of unlabelled data, deep learning has been applied in many fields including the medical field. In this article, the most noteworthy applications of deep learning are presented shortly and positively, they are image recognition, automatic speech recognition, natural language processing, drug discovery and toxicology, customer relationship management, recommendation systems and bioinformatics. The report concluded that these applications have a significant and vital role in all areas of life.

Technological improvements lead big data producing, processing and storing systems. These systems must contain extraordinary capabilities to overcome complexity of the big data. Therefore, the methodologies utilized for data analysis have been evolved due to the increase in importance of extracting information from big data. Healthcare systems are important systems dealing with big data analysis. Deep learning is the most applied data analysis method. It becomes one of the most popular and up-to-date artificial neural network types with deep representation ability. Another powerful ability of deep learning is providing feature learning through convolutional neural networks. Deep learning has wide implementation areas in medical applications from diagnosis to treatment. Various deep learning methods are applied to the biomedical problems. In many applications, deep learning solutions are modified in accordance with the requirements of the problems. Through this chapter the most popul...

moch chamadani

Loading Preview

Sorry, preview is currently unavailable. You can download the paper by clicking the button above.

RELATED PAPERS

Nilay Ganatra

Journal of Al-Qadisiyah for Computer Science and Mathematics

Raidah S. Khudeyer

Pattern Recognition and Image Analysis

Aliaksandr Kroshchanka

Manuel Valenzuela-Rendón

Debasmita Bhoumik

ritesh patra

Electronics

Vijayan Asari

Oduntan Adeola

oduntan haydex

International Journal of Electrical and Computer Engineering (IJECE)

Priyanka Patel

IEEE Access

Ajay Shrestha

International Journal of Artificial Intelligence & Mathematical Sciences

Arifa Bhutto

ResearchGate

Fabio Grazioso

NISA SAPUTRA

International Journal of Engineering Research and Technology (IJERT)

IJERT Journal

Applied Journal of Physical Science Volume 4(4), pages 42-50, December 2022 Article Number: 1378CAD82 ISSN: 2756-6684 https://doi.org/10.31248/AJPS2022.085 https://integrityresjournals.org/journal/AJPS

Ayobami Ayeni

Deep Learning and Edge Computing Solutions for High Performance Computing

lakshmi haritha

The Research Publication

Suban Ravichandran

Journal of Soft Computing Paradigm

Prof. Dr. Subarna Shakya

IAEME Publication

Srivatsa Sarat Kumar Sarvepalli

FRANK KATAKA BANASEKA , Ben Bright Benuwa , Benjamin Ghansah (P.hD.) , Dickson Keddy

Applied Deep Learning - Part 2: Real World Case Studies

Arden Dertat

Towards Data Science

Welcome to Part 2 of Applied Deep Learning series. Part 1 was a hands-on introduction to Artificial Neural Networks, covering both the theory and application with a lot of code examples and visualization. Now comes the cool part, end-to-end application of deep learning to real-world datasets. We will cover the 3 most commonly encountered problems as case studies: binary classification, multiclass classification and regression.

• Case Study: Binary Classification 1.1) Data Visualization & Preprocessing 1.2) Logistic Regression Model 1.3) ANN Model 1.4) Visualization of Deep ANN
• Case Study: MultiClass Classification 2.1) Data Visualization & Preprocessing 2.2) Softmax Regression Model 2.3) ANN Model 2.4) Cross Validation
• Case Study: Regression 3.1) Data Visualization & Preprocessing 3.2) Linear Regression Model 3.3) ANN Model

The code for this article is available here as a Jupyter notebook, feel free to download and try it out yourself.

1. Case Study: Binary Classification

We will be using the Human Resources Analytics dataset on Kaggle. We’re trying to predict whether an employee will leave based on various features such as number of projects they worked on, time spent at the company, last performance review, salary etc. The dataset has around 15,000 rows and 9 columns. The column we’re trying to predict is called “left”. It’s a binary column with 0/1 values. The label 1 means that the employee has left.

1.1) Data Visualization & Preprocessing

First things first, let’s perform some data visualization and preprocessing before jumping straight into building the model. This part is crucial, since we need to know what type of features we are dealing with. For every ML task, we at least need to answer the following questions:

• What type of features do we have: real valued, categorical, or both?
• Do any of the features need normalization?
• Do we have null values?
• What is the label distribution, are the classes imbalanced?
• Is there a correlation between the features?

The jupyter notebook contains the detailed analysis. In summary, there are both real and categorical features. There are no null values, but some features need normalization. 76% percent of the examples are labeled as 0, meaning the employee didn’t leave.

Let’s check the correlation of the features with the labels (the column named “left”). We will use the seaborn package for the correlation plot.

In this plot, positive values represent correlation and negative values represent inverse correlation with the label. Of course “left” has perfect correlation with itself, you can ignore that. Other than that only one feature has a strong signal, which is the “satisfaction_level”, inversely correlated with whether the employee has left. Which makes sense.

Now let’s look at the pairwise correlation of all features with one another.

We see that “average_monthly_hours” is positively correlated with “number_project”, which again makes sense. The more projects a person is involved with, the more hours of work they need to put in.

Now let’s look at the distribution of feature values. By inspecting the histograms of features we can see which ones need normalization. What’s the motivation behind this? What does normalization mean and why is it needed? Most ML algorithms perform better if the real valued features are scaled to be in a predefined range, for example [0, 1]. This is particularly important for deep neural networks. If the input features consist of large values, deep nets really struggle to learn. The reason is that as the data flows through the layers, with all the multiplications and additions, it gets large very quickly and this negatively affects the optimization process by saturating non-linearities. We will see the detailed demonstration of this in another article, for now we need to pay attention to feature values to be small numbers.

Looking at feature histograms, we need to normalize 3 of the features: average_monthly_hours, number_project, and time_spend_company. All other features are within [0, 1] so we can leave them alone.

Scikit-learn has several normalization methods, what we will use is StandardScaler . It individually scales the features such that they have zero mean and unit variance, so they all belong to a standard Normal(0, 1) distribution. Note that this doesn’t change the ordering of the feature values, it just changes the scale. It’s a simple yet extremely important trick.

The data we loaded is in a pandas DataFrame. Pandas is an extremely popular package to deal with tabular data, especially in an interactive environment like jupyter notebooks. DataFrame is the most commonly used data structure of pandas that acts as a container for our data, and exposes several built-in functions to make our life easier (check out the notebook for more details). In the code snippet below, df is the DataFrame for our data.

Scikit-learn API is very well designed and contains 4 very commonly used methods. Predictors are ML models like Logistic Regression, and transformers are data manipulators like Standard Scaler.

• fit : For predictors performs training on the given input. For transformers computes the statistics like mean and standard deviation of the input to be used later.
• transform : For transformers manipulates the input data using the stats learned by the fit function. We run the transform method after fit since there’s a dependency. Predictors don’t support this method.
• fit_transform : Performs fit + transform in a single call efficiently. For transformers, computes the stats of the input and performs the transformation. It’s very a commonly used method with transformers. For predictors, trains the model and performs prediction on the given input.
• predict: As its name suggests, for predictors performs the prediction task using the model trained with the fit method. Very commonly used with predictors. Transformers don’t support this method.

Now that we have scaled the real-valued features to be in a desirable range, let’s deal with the categorical features. We need to convert categorical data to one-hot representation. For example the salary column contains 3 unique string values: low, medium and high. After one-hot conversion we will have 3 new binary columns: salary_low, salary_medium and salary_high. For a given example, only one of them will have the value 1, the others will be 0. We will then drop the original salary column because we don’t need it anymore.

The one-hot conversion is performed by the get_dummies of pandas. We could have also used the OneHotEncoder in scikit-learn, they both get the job done. Since our data is already is in a pandas dataframe, get_dummies is easier. It also automatically perform the renaming of the features.

Now comes the final part of creating the training and test data. The model will perform learning on the training set and be evaluated on the held-out test set. Scikit-learn has a convenient train_test_split function. We only need to specify the fraction of the test set, in our case 30%. But first we convert our data from pandas dataframe to numpy array using the values attribute of the dataframe.

1.2) Logistic Regression Model

Now that we’re done with the data preprocessing and train/test set generation, here comes the fun part, training the model. We first start with a simple model, Logistic Regression (LR). We will then train a deep ANN and compare the results to LR.

After the first article, building the model should be very familiar.

We get 79% training accuracy. This is actually pretty bad, because above we saw that 76% of the labels were 0. So the most naive classifier which always outputs 0 regardless of the input would get 76% accuracy, and we’re not doing much better than that. This means our data is not linearly separable, just like the examples we saw in the first article, and we need a more complex model.

Above chart depicts the training loss and accuracy. But more importantly we’re interested in the metrics of the test set. Metrics in the training set might be misleading since the model is already trained on it, we want to check how the model performs on an held-out test set. Test accuracy is 78%, slightly lower than training accuracy. Test accuracy of ML models are almost always less than training, because the test data is unseen to the model during the training process. Looking at the classification report, we see that only 60% of the examples belonging to class 1 are classified correctly. Pretty bad performance. The confusion matrix also doesn’t look promising showing a lot of misclassified examples.

1.3) ANN Model

Now let’s build a deep neural network for binary classification. This model will be much more powerful, and will be able to model non-linear relationships.

The model building process is again very familiar. We have 2 hidden layers with 64 and 16 nodes with tanh activation function. The output layer uses the sigmoid activation since it’s a binary classification problem. We use the Adam optimizer with learning rate set to 0.01.

This time we achieve 97.7% training accuracy, pretty good.

Let’s compare the LR and ANN models. The ANN model is much superior, having a lower loss and a higher accuracy.

And for completeness here’s the classification report and confusion matrix of the ANN model on the test set. We achieve 97% accuracy, compared to 78% of the LR model. We still misclassify 147 examples out of 4500.

We can further improve the performance of the ANN by doing the following:

• Train the model for longer (increase the number of epochs).
• Hyperparamter tuning: change the learning rate, use a different optimizer than Adam (RMSprop for example), use another activation function than tanh (can be relu).
• Increase the number of nodes per layer: Instead of 64–16–1 we can do 128–64–1.
• Increase the number of layers: We can do 128–64–32–16–1.

One important caveat though, as we make the model more powerful, the training loss will likely decrease and accuracy will increase. But we will run into the risk of overfitting. Meaning the complex model will perform worse on the test set compared to a simpler model, even though the training metrics of the complex model is better. We will talk more about overfitting in another article, but this is very important to keep in mind. That’s why we don’t go crazy with number of layers and nodes per layer. The simplest model that gets the job done is sufficient.

1.4) Visualization of Deep ANN

In the previous article we learned that each layer of the ANN performs a non-linear transformation of the input from one vector space to another. By doing this we are projecting our input data to a new space where the classes are separable from each other via a complex decision boundary.

Let’s visually demonstrate this. Our input data after the initial data preprocessing we did above is 20 dimensional. For visualization purposes let’s project it to 2D. Remember that having k nodes in a layer means that this layer transforms its input such that the output is a k-dimensional vector. The ANN we trained above had two hidden layers with 64 and 16 nodes. Then we need a new layer with 2 nodes in order to project our data to a 2D space. So we add this new layer just before the output node. The rest is completely untouched.

Here’s the resulting projection of our input data from 20D to 2D space. The decision boundary corresponds to the last layer of the ANN. The ANN was able to separate out the classes pretty well, with some misclassifications. A lot of data points overlap in 2D so we can’t see them all, for reference the model misclassifies around 160 points out of 4500 (96% accuracy). We aren’t concerned about accuracy with this model anyway, we are interested in the projection of a high-dimensional input to 2D. This is a neat little trick to visually demonstrate the result of the projections performed by the ANN.

A more principled visualization approach would be using t-SNE , which is a dimensionality reduction technique for visualizing high-dimensional data. Details available here .

2. Case Study: MultiClass Classification

We will now perform multiclass classification on the famous Iris dataset . It contains 3 classes of flowers with 50 examples each. There are a total of 4 features. So it’s pretty small, but very popular. The data looks as follows.

2.1) Data Visualization & Preprocessing

This part is easier now since we only have 4 real-valued features. Again we normalize the features to be between [0, 1]. The feature values are small and we could get away without any normalization, but it’s a good habit to do so, and there’s no harm doing it anyway.

Pairplot is a cool visualization technique if we have a small number of features. We can see the pairwise distribution of features colored by the class (the column “label” in the dataset). We use the seaborn package, pretty simple for an informative plot.

We can see that the “setosa” class is easily separable but “versicolor” and “virginica” are more intermingled.

2.2) Softmax Regression Model

We will now train a Softmax Regression (SR) model to predict the labels. The previous article contains a detailed explanation, and building the model is again very similar to above. The main difference is that we’re using softmax activation and categorical_crossentropy as the loss.

The SR model already achieves 97% training accuracy and a minimal loss, pretty good already.

2.3) ANN Model

Now let’s build our ANN model. We add 2 hidden layers with 32 and 16 nodes. Notice that we also change the activation function of these layers to relu instead of tanh. We will explore various activation functions and their differences in another tutorial, but relu is arguably the most popular one. Then why we haven’t been using it? Well, only because the decision boundary plots looked prettier with tanh activation. Seriously, no other reason.

This time we get 100% traning accuracy.

And for the record, test accuracy of both SR and ANN models are 100%. This is a pretty small dataset, so these results are not surprising. Not all problems need a deep neural net to get good results. For this problem it’s probably an overkill, since a linear model works just as fine.

2.4) Cross Validation

With a small sample size like our current situation, it’s especially important to perform cross validation to get a better estimate on accuracy. With k-fold cross validation we split the dataset into k disjoint parts, use k-1 parts for training and the other 1 part for testing. This way every example appears in both training and test sets. We then average out the model’s performance in all k runs and get a better low-variance estimation of the model accuracy.

Usually while training deep learning models we don’t perform k-fold cross validation. Because the training takes a long time, and training the model k times from scratch is not feasible. But since our dataset is small it’s a good candidate to try on.

Here’s the plot for 5-fold cross validation accuracy for both models. The deep model is doing slightly better, has a higher accuracy and lower variance. In the figure the accuracies sometimes seem to be above 100% but that’s an artifact of smoothing the curves. Max accuracy we get is 100%.

3. Case Study: Regression

We’ll now work on a regression problem, predicting a real-valued output instead of discreet class memberships. We will be using the house sales dataset from King County, WA on Kaggle. There are around 21,000 rows with 20 features. The value we’re trying to predict is a floating point number labeled as “price”.

3.1) Data Visualization & Preprocessing

First let’s take a look at feature distributions

You know the drill by now, we need to do feature normalization and categorization. For example squarefoot related features definitely need to be normalized since the values range in thousands, and features like zipcode need to be categorized.

We also have a new type of preprocessing to do, bucketization . For example the feature which contains the year the house was built (yr_built), ranges from 1900 to 2015. We can certainly categorize it with every year belonging to a distinct category, but then it would be pretty sparse. We would get more signal if we bucketized this feature without losing much information. For example if we use 10 year buckets, years between [1950, 1959] would be collapsed together. It would probably be sufficient to know that this house was built in 1950s instead of 1958 exactly.

Other features that would benefit from bucketizing are latitude and longitude of the house. The exact coordinate doesn’t matter that much, we can round the coordinate to the nearest kilometer. This way the feature values will be more dense and informative. There’s no hard and set rule to which ranges to use in bucketization, they’re mostly decided by trial and error.

One final transformation we need to do is for the price of the house, the value we’re trying to predict. Currently its value ranges from $75K to $7.7M. A model trying to predict in such a large scale and variance would be very unstable. So we normalize that as well. Feel free to check the code for the details.

After all the transformations we go from 20 to 165 features. Let’s check the correlation of each feature with price.

The most correlated feature is the square footage, which is expected, bigger houses are usually more expensive. Looking at the list the features make sense. Some zipcodes have high correlation with price, for example 98039 which corresponds to Medina, that’s where Bill Gates lives and it’s one of the most expensive neighborhoods in the country. There’s another zipcode 98004 which is more correlated corresponding to Bellevue. There are a lot of high-rises and tech offices there, which has been driving up the prices a lot lately. I used to live in that neighborhood, but then it got too boring and expensive so I moved :)

3.2) Linear Regression Model

This is the first time we’re building a regression model. Just like Logistic Regression is the simplest model we try first in a classification problem, Linear Regression is the one we start with in a regression problem.

Remember that the equation for logistic regression is y=f(xW) where f is the sigmoid function. Linear regression is simply y=xW, that’s it no activation function. I’ve again omitted the bias terms for simplicity. With the biases they become y=f(xW+b) and y=xW+b respectively.

As a reminder here is how we built the logistic regression (LR) model

And here’s the code for the linear regression model (LinR)

There are 3 main differences:

• LR uses a sigmoid activation function, where LinR has no activation.
• LR uses binary_crossentropy loss function, where LinR uses mean_squared_error.
• LR also reports the accuracy, but accuracy is not an applicable metric to a regression problem, since the output is a floating point number instead of a class membership.

The most important change is the loss function mean_squared_error (MSE). MSE is a standard loss function used for regression. The formula is pretty simple:

Where y is the true value, ŷ is the predicted and n is the number of examples.

We’re also passing a new validation_split argument to the fit function. It specifies the fraction of training data to use as held-out validation set during training, in our case it’s 20%. Using a validation set we can see whether we’re overfitting during training. Don’t confuse the validation set with the test set though. Test set is entirely separate and it doesn’t get exposed to the model at all during training.

The loss decreases in the first couple of epochs and then stabilizes. We’re likely underfitting meaning our model doesn’t have enough capacity, and we need a more compex model.

And these are the top 10 features with the highest weights. Categorical zipcode features are highly dominant.

3.3) ANN Model

And finally let’s build an ANN for regression. In the previous examples, going from a linear model to a deep model just involved adding new layers with non-linear activation functions. It will be the same this time as well.

We added new layers with relu activation. The loss plot looks interesting now. The training error loss still seems to be decreasing, but the validation error starts increasing after the 5th epoch. We’re clearly overfitting . The ANN is memorizing the training data, and this is reducing its ability to generalized on the validation set.

The solution? There are a couple of methods for tackling overfitting in deep neural nets. Most of the methods rely on constraining the capacity of the model. This can be achieved for example by restricting weights, sharing weights or even stopping training before the training loss plateaus. We will use the last one in the interest of brevity and discuss the rest in another article. This means we will simply halt training as soon as the validation loss stops improving. This is called early stopping , and it’s pretty easy to implement with Keras. We just need to modify the call to the fit function as follows. If there’s no improvement for validation loss for 2 epochs, we stop training.

The validation loss stopped improving at the 5th epoch, the model waited for 2 more epochs and finished training at epoch 7.

The training loss with the LinR model was 0.158, the loss for ANN is 0.086. We got 83% improvement over the linear model, pretty good. And comparing the loss on the test set, LinR model got 0.191 compared to 0.127 of ANN which is 32% improvement. The figure below compares the training loss of LinR vs ANN.

Let’s make a final comparison, dollar value difference between the model predicted price and actual price. The most naive model which always predicts the average price of the training set ($540K), is off by $229K on the test set, pretty bad. Linear Regression model is off by $87K vs $68K for the Deep ANN . The ANN is 21% better than LinR.

4) Conclusion

I hope this was an informative article. I tried to demonstrate a step-by-step application of deep learning on 3 common machine learning problems with real-life datasets.

Data preprocessing and visualization was covered in detail. Even though they’re not the fun part of solving an ML problem and mostly overlooked, they’re extremely important. Just like Part 1 we first tackled the problem with a simple model, and then used a deep neural net to get better results.

The entire code for this article is available here if you want to hack on it yourself. If you have any feedback feel free to reach out to me on twitter .

Written by Arden Dertat

ML Engineer @ Netflix. Photography and travel enthusiast.

Text to speech

IEEE Account

• Change Username/Password
• Update Address

Purchase Details

• Payment Options
• Order History
• View Purchased Documents

Profile Information

• Communications Preferences
• Profession and Education
• Technical Interests
• US & Canada: +1 800 678 4333
• Worldwide: +1 732 981 0060
• Contact & Support
• About IEEE Xplore
• Accessibility
• Terms of Use
• Nondiscrimination Policy
• Privacy & Opting Out of Cookies

A not-for-profit organization, IEEE is the world's largest technical professional organization dedicated to advancing technology for the benefit of humanity. © Copyright 2024 IEEE - All rights reserved. Use of this web site signifies your agreement to the terms and conditions.

• Open access
• Published: 12 November 2020

Deep learning accelerators: a case study with MAESTRO

• Hamidreza Bolhasani   ORCID: orcid.org/0000-0003-0698-6141 1 &
• Somayyeh Jafarali Jassbi 1  

Journal of Big Data volume  7 , Article number:  100 ( 2020 ) Cite this article

8120 Accesses

6 Citations

25 Altmetric

Metrics details

In recent years, deep learning has become one of the most important topics in computer sciences. Deep learning is a growing trend in the edge of technology and its applications are now seen in many aspects of our life such as object detection, speech recognition, natural language processing, etc. Currently, almost all major sciences and technologies are benefiting from the advantages of deep learning such as high accuracy, speed and flexibility. Therefore, any efforts in improving performance of related techniques is valuable. Deep learning accelerators are considered as hardware architecture, which are designed and optimized for increasing speed, efficiency and accuracy of computers that are running deep learning algorithms. In this paper, after reviewing some backgrounds on deep learning, a well-known accelerator architecture named MAERI (Multiply-Accumulate Engine with Reconfigurable interconnects) is investigated. Performance of a deep learning task is measured and compared in two different data flow strategies: NLR (No Local Reuse) and NVDLA (NVIDIA Deep Learning Accelerator), using an open source tool called MAESTRO (Modeling Accelerator Efficiency via Spatio-Temporal Resource Occupancy). Measured performance indicators of novel optimized architecture, NVDLA shows higher L1 and L2 computation reuse, and lower total runtime (cycles) in comparison to the other one.

Introduction

The main idea of neural networks (NN) is based on biological neural system structure, which consists of several connected elements named neurons [ 1 ]. In biological systems, neurons get signals from dendrites and pass them to the next neurons via axon as shown in Fig.  1 .

Typical biological neurons [ 20 ]

Neural networks are made up of artificial neurons for handling brain tasks like learning, recognition and optimization. In this structure, the nodes are neurons, links can be considered as synapses and biases as activation thresholds [ 2 ]. Each layer extracts some information related to the features and forwards them with a weight to the next layer. Output is the sum of all these information gains multiplied by their related weights. Figure  2 represents a simple artificial neural network structure.

Simple artificial neural network structure

Deep neural networks are complex artificial neural networks with more than two layers. Nowadays, these networks are widely used for several scientific and industrial purposes such as visual object detection, segmentation, image classification, speech recognition, natural language processing, genomics, drug discovery, and many other areas [ 3 ].

Deep learning is a new subset of machine learning including algorithms that are used for learning concepts in different levels, utilizing artificial neural networks [ 4 ].

As Fig.  3 shows, if each neuron and its weight are represented by X i and W i j respectively, the output result (Y j ) would be:

A typical deep neural network structure

where \(\sigma\) is the activation function. A popular function that is used for activation in deep neural networks is ReLU (Rectified Linear Unit) function, which is defined in Eq. ( 2 ).

Leaky ReLU, tanhh and Sigmoid functions are some other activation functions with less frequent usage [ 5 ].

As shown in Fig.  4 , each layer of a deep neural network’s role is to extract some features and send them to the next layer with its corresponding weight. For example, in the first layer, color properties (green, red blue) are gained; in the next layer, edge of objects are determined and so on.

Deep learning setup for object detection [ 21 ]

Convolutional neural networks are a type of deep neural networks that is mostly used for recognition, mining and synthesis applications like face detection, handwritting recognition and natural language processing [ 6 ]. Since parallel computations is an unavoidable part of CNNs, several efforts and research works have been done for designing an optimized hardware for it. As a result, many application-specific integrated circuits (ASICs) as hardware accelerators have been introduced and evaluated in the recent decade [ 7 ]. In the next section, some of the most successful and impressive works related to CNN accelerators are introduced.

Related works

Tianshi et al. [ 8 ] proposed DianNao as a hardware accelerator for large-scale convolutional neural networks (CNNs) and deep neural networks (DNNs). The main focus of the suggested model is on the memory structure to be optimized for big neural network computations. The experimental results showed speedup in computation and reduction of overhead in performance and energy. This research also demonstrated that the accelerator can be implemented in very small area in order of 3 mm 2 and 485 mW power.

Zidong et al. [ 9 ] suggested ShiDianNao as a CNN accelerator for image processing close to a CMOS or CCD sensor. The performance and energy of this architecture is compared to CPU, GPU and DainNao, which has been discussed in previous work [ 8 ]. Utilizing SRAM instead of DRAM made it 60 times more enery effiecent than DianNao. It is also 50×, 30× and 1.87× faster than a mainstream CPU, GPU and DianNao, with just 65 nm usage area and 320 mW power.

Wenyan et al. [ 6 ] offered a flexible dataflow accelerator for convolutional neural networks called FlexFlow. Working on different types of parallelism is the substantial contribution of this model. Results of the tests showed 2–10 × performance speedup and 2.5–10 × power efficiency in comparison with three investigated baseline architectures.

Eyriss is a spatial architecture for energy efficient data flows for CNNs which presented by Yu-Hsin et al. [ 10 ]. This hardware model is based on a dataflow named row stationary (RS). This dataflow minimizes energy consumption by reusing computation of filter weights. The proposed RS dataflow is investigated on AlexNet CNN configuration, which proved energy efficiency improvement.

Morph is a flexible accelerator for 3D CNN-based video processing that offered by Katrik et al. [ 7 ]. Since the previous work and proposed architectures didn’t specificly focus on video processing, this model can be considered as a novelty in this area. Comparison of energy consumption in this architecture with previous idea, Eyriss [ 10 ] showed a high level of reduction that means energy saving. The main reason of this improvement is effective data reuse which reduces the access to higher level buffers and high cost off-cheap memory.

Michael et al. [ 11 ] described Buffets that is an efficient and composable accelerator and independent of any particular design. Through this research, explicit decoupled data orchestration (EDDO) is introduced which allows evaluation of energy efficiency in acceleators. Result of this work showed that with a smaller usage area, higher energy efficiency and lower control overhead is acquired.

Deep learning applications

Deep learning has a wide range of applications in recognition, classification and prediction, and since it tends to work like the human brain and consequently does the human jobs in a more accurate and low cost manner, its usage is dramatically increasing. More than 100 papers published from 2015 to 2020, helped categorize the main applications as below:

Computer vision

Translation

Health monitoring

Disease prediction

Medical image analysis

Drug discovery

Biomedicine

Bioinformatics

Smart clothing

Personal health advisors

Pixel restoration for photos

Sound restoration in videos

Describing photos

Handwriting recognition

Predicting natural disasters

Cyber physical security systems [ 12 ]

Intelligent transportation systems [ 13 ]

Computed tomography image reconstruction [ 14 ]

As mentioned previously, artificial intelligence and deep learning applications are growing drastically, but they have high complexity computation, energy consumption, costs and memory bandwidth. All these reasons were major motivations for developing deep learning accelerators (DLA) [ 15 ]. A DLA is a hardware architecture that is specially designed and optimized for deep learning purposes. Recent DLA architectures (e.g. OpenCL) have mainly focused on maximizing computation reuse and minimizing memory bandwidth, which led to higher speed and performance [ 16 ].

Generally, most of the accelerators support just fixed data flow and are not reconfigurable, but for doing huge deployments, they need to be programmable. Hyoukjun et al. [ 15 ] proposed a novel architecture named MAERI (Multiply-Accumulate Engine with Reconfigurable Interconnects), which is reconfigurable and employs ART (Augmented Reduction Tree) which showed 8 ~ 459% better utilization for different data flows over a strict network-on-chip (NoC) fabric. Figure  5 shows the overall structure of MAERI DLA.

MAERI micro architecture [ 15 ]

In another research, Hyoukjun et al. offered a framework called “MAESTRO” (Modeling Accelerator Efficiency via Spatio-Temporal Resource Occupancy) for predicting energy performance and efficiency in DLAs [ 17 ]. MAESTRO is an open-source tool that is capable of computing many NoC parameters for a proposed accelerator and related data flow such as maximum performance (roofline throughput), compute runtime, total runtime, NoC analysis, L1 to L2 NoC bandwidth, L2 to L1 bandwidth analysis, buffer analysis, L1 and L2 computation reuse, L1 and L2 weight reuse, L1 and L2 input reuse and so on. The topology, tool flow and relationship between each of its blocks of this framework are presented in Fig.  6 .

MAESTRO topology [ 15 ]

Results and discussion

In this paper, we used MAESTRO to investigate buffer, NoC, and performance parameters of a DLA in comparison to a classical architecture for a specific deep learning data flow. For running MAESTRO and getting the related analysis, some parameters should be configured, as follows:

LayerFile: Including the information related to the layers of neural network.

DataFlow File: Information related to data flow.

Vector Width: Width of the vectors.

NoCBand width: Bandwidth of NoC.

Multicast Supported: This logical indictor (True/False) is for defining that the NoC supports multicast or not.

NumAverageHopsinNoC: Average number of hops in the NoC.

NumPEs: Number of processing elements.

For the simulation of this paper, we configured the mentioned parameters as presented in Table 1 .

As presented in Table 1 , we have selected Vgg16_conv11 as LayerFile, which is a convolutional neural network that has proposed by K. Simonyan and A. Zisserman. This deep convolutional network model was offered for image recognition with 92.7% accuracy on ImageNet dataset [ 18 ].

Two different data flow strategies are investigated and compared in this study: NLR and NVDLA. NLR stands for “No Local Reuse” which expresses its specific strategy and NVDLA is a novel DLA designed by NVIDIA Co [ 19 ].

Other parameters such as vector width, NoC bandwidth, multicast support capability, average numbers of hops and numbers of processing elements in NoC have been selected based on a real hardware condition.

Simulation results demonstrated that NVDLA has better performance, runtime, higher computation reuse and lower memory bandwidth in comparison to NLR as presented in Table 2 and Figs. 7 , 8 , and 9 .

Comparing L1 Weight and Input Reuse

Comparing L2 Weight and Input Reuse

Total Runtime comparison

Artificial intelligence, machine learning and deep learning are growing trends affecting our lives in almost all aspects of human’s life. These technologies make our life easier by assigning routine tasks of human resources to the machines that are much more accurate and fast. Therefore, any effort for optimizing performance, speed, and accuracy of these technologies is valuable. In this research, we focused on performance improvements of the hardware that are used for deep learning purposes named deep learning accelerators. Investigating recent researches conducted on these hardware accelerators shows that they can optimize costs, energy consumption, run time about 8–459% based on MAERI’s investigation by minimizing memory bandwidth and maximizing computation reuse. Utilizing an open source tool named MAESTRO, we compared buffer, NoC and performance parameters of NLR and NVDLA data flows. Results showed higher computation reuse for both L1 and L2 of the NVDLA data flow which is designed and optimized for deep learning purposes and studied as deep leraning accelerator in this study. The results showed that the customized hardware accelartor for deep learning (NVDLA) had much shorter total runtime in comparison with NLR.

Availability of data and materials

Abbreviations.

Multiply-accumulate engine with reconfigurable interconnects

No local reuse

NVIDIA deep learning accelerator

Modeling accelerator efficiency via spatio-temporal resource occupancy

Rectified linear unit

• Deep learning accelerator

Neural network

Convolutional neural network

Deep neural network

Row stationary

Application-specific integrated circuits

Augmented reduction tree

Network on chip

L1 read sum

L1 write sum

L2 read sum

L2 write sum

Jurgen S. Deep learning in neural networks: an overview. Neural Netw. 2015;61:85–117.

Article   Google Scholar  

Muller B, Reinhardt J, Strickland MT. Neural networks: an introduction. Berlin: Springer; 2012. p. 14–5.

MATH   Google Scholar  

Yann L, Yoshua B, Geoffrey H. Deep learning. Nature. 2015;521:436–44.

Li D, Dong Y. Deep learning: methods and applications. Found Trends Signal Process. 2014;7:3–4.

MathSciNet   Google Scholar  

Jianqing F, Cong M, Yiqiao Z. A selective overview of deep learning. arXiv:1904.05526[stat.ML]. 2019.

Wenyan L, et al. FlexFlow: a flexible dataflow accelerator architecture for convolutional neural networks. In: IEEE ınternational symposium on high performance computer architecture. 2017.

Katrik H, et al. Morph: flexible acceleration for 3D CNN-based Video Understanding. In 51st annual IEEE/ACM international symposium on microarchitecture (MICRO). 2018.

Tianshi R, et al. DianNao: a small-footprint high-throughput accelerator for ubiquitous machine-learning. In: ACM SIGARCH Computer Architecture News; 2014.

Zidong D, et al. ShiDianNao: Shifting Vision Processing Closer to the Sensor. In: ACM/IEEE 42nd annual ınternational symposium on computer architecture (ISCA). 2015.

Chen Y-H, Emer J, Sze V. Eyeriss: A spatial architecture for energy-efficient dataflow for convolutional neural networks. ACM SIGARCH Computer Achitect News. 2016;44:367–79.

Michael P, et al. Buffets: an efficient and composable storage idiom for explicit decoupled data orchestration. In: ASPLOS '19: proceedings of the twenty-fourth ınternational conference on architectural support for programming languages and operating systems. 2019, P 137.

Xia X, Marcin W, Fan X, Damasevicius R, Li Y. Multi-sink distributed power control algorithm for Cyber-physical-systems in coal mine tunnels. Comput Netw. 2019;161:210–9.

Song H, Li W, Shen P, Vasilakos A. Gradient-driven parking navigation using a continuous information potential field based on wireless sensor network. Inf Sci. 2017;408(2):100–14.

Google Scholar  

Bin Z, Dawid P, Marcin W. A regional adaptive variational PDE model for computed tomography image reconstruction. Pattern Recogn. 2019;92:64–81. https://doi.org/10.1016/j.patcog.2019.03.009 .

Hyoukjun K, Ananda S, Tushar K. MAERI: enabaling flexible dataflow mapping over DNN accelerators via reconfigurable interconnetcs. In: ASPLOS ’18, Proceedings of the twenty-third international conference on architectural support for programming languages and operating systems.

Uktu A, Shane O, Davor C, Andrew C. L, and Gordon R. C. An OpenCL deep learning accelerator on Arria 10. In: FPGA ’17, Proceedings of the 2017 ACM/SIGDA international symposium on field programmable gate arrays, pp. 55–64.

Hyoukjun K, Michael P, Tushar K. MAESTRO: an open-source infrastructure for modeling dataflows within deep learning accelerators. arXiv:1805.02566. 2018.

Karen S, Andrew Z. Very deep convolutional network for large-scale image recognition. arXiv:1409.1556. 2015

NVDLA Deep Learning Accelerator, https://nvdla.org . 2017.

George SE. The anatomy and physiology of the human stress response. A clinical guide to the treatment of the human stress responses. Berlin: Springer, pp 19–56.

Christian S, Alexander T, Dumitru E. Deep neural networks for object detection. Advances in neural information processing systems 26, NIPS 2013.

Download references

Acknowledgements

Not applicable.

Author information

Authors and affiliations.

Department of Computer Engineering, Science and Research Branch, Islamic Azad University, Tehran, Iran

Hamidreza Bolhasani & Somayyeh Jafarali Jassbi

You can also search for this author in PubMed   Google Scholar

Contributions

Investigating deep learning accelerators functionality. Analyzing a deep learning accelerator’s architecture. Performance measurement of NVIDIA deep learning accelerator as a case study. Higher computation reuse and lower total runtime for the studied deep learning accelerator in comparison with non-optimized architecture.

Corresponding author

Correspondence to Hamidreza Bolhasani .

Ethics declarations

Competing interests.

Evaluating a deep learning accelerator’s performance.

Additional information

Publisher's note.

Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Rights and permissions

Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article's Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article's Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/ .

Reprints and permissions

About this article

Cite this article.

Bolhasani, H., Jassbi, S.J. Deep learning accelerators: a case study with MAESTRO. J Big Data 7 , 100 (2020). https://doi.org/10.1186/s40537-020-00377-8

Download citation

Received : 13 May 2020

Accepted : 01 November 2020

Published : 12 November 2020

DOI : https://doi.org/10.1186/s40537-020-00377-8

Share this article

Anyone you share the following link with will be able to read this content:

Sorry, a shareable link is not currently available for this article.

Provided by the Springer Nature SharedIt content-sharing initiative

• Deep learning
• Convolutional neural networks
• Deep neural networks
• Hardware accelerator

Deep Learning for Recommender Systems: A Netflix Case Study

• Harald Steck Netflix
• Linas Baltrunas Netflix
• Ehtsham Elahi Netflix
• Dawen Liang Netflix
• Yves Raimond Netflix
• Justin Basilico Netflix

Deep learning has profoundly impacted many areas of machine learning. However, it took a while for its impact to be felt in the field of recommender systems. In this article, we outline some of the challenges encountered and lessons learned in using deep learning for recommender systems at Netflix. We first provide an overview of the various recommendation tasks on the Netflix service. We found that different model architectures excel at different tasks. Even though many deep-learning models can be understood as extensions of existing (simple) recommendation algorithms, we initially did not observe significant improvements in performance over well-tuned non-deep-learning approaches. Only when we added numerous features of heterogeneous types to the input data, deep-learning models did start to shine in our setting. We also observed that deep-learning methods can exacerbate the problem of offline–online metric (mis-)alignment. After addressing these challenges, deep learning has ultimately resulted in large improvements to our recommendations as measured by both offline and online metrics. On the practical side, integrating deep-learning toolboxes in our system has made it faster and easier to implement and experiment with both deep-learning and non-deep-learning approaches for various recommendation tasks. We conclude this article by summarizing our take-aways that may generalize to other applications beyond Netflix.

How to Cite

• Endnote/Zotero/Mendeley (RIS)
• The author(s) warrants that they are the sole author and owner of the copyright in the above article/paper, except for those portions shown to be in quotations; that the article/paper is original throughout; and that the undersigned right to make the grants set forth above is complete and unencumbered.
• The author(s) agree that if anyone brings any claim or action alleging facts that, if true, constitute a breach of any of the foregoing warranties, the author(s) will hold harmless and indemnify AAAI, their grantees, their licensees, and their distributors against any liability, whether under judgment, decree, or compromise, and any legal fees and expenses arising out of that claim or actions, and the undersigned will cooperate fully in any defense AAAI may make to such claim or action. Moreover, the undersigned agrees to cooperate in any claim or other action seeking to protect or enforce any right the undersigned has granted to AAAI in the article/paper. If any such claim or action fails because of facts that constitute a breach of any of the foregoing warranties, the undersigned agrees to reimburse whomever brings such claim or action for expenses and attorneys’ fees incurred therein.
• Author(s) retain all proprietary rights other than copyright (such as patent rights).
• Author(s) may make personal reuse of all or portions of the above article/paper in other works of their own authorship.
• Author(s) may reproduce, or have reproduced, their article/paper for the author’s personal use, or for company use provided that original work is property cited, and that the copies are not used in a way that implies AAAI endorsement of a product or service of an employer, and that the copies per se are not offered for sale. The foregoing right shall not permit the posting of the article/paper in electronic or digital form on any computer network, except by the author or the author’s employer, and then only on the author’s or the employer’s own web page or ftp site. Such web page or ftp site, in addition to the aforementioned requirements of this Paragraph, must provide an electronic reference or link back to the AAAI electronic server, and shall not post other AAAI copyrighted materials not of the author’s or the employer’s creation (including tables of contents with links to other papers) without AAAI’s written permission.
• Author(s) may make limited distribution of all or portions of their article/paper prior to publication.
• In the case of work performed under U.S. Government contract, AAAI grants the U.S. Government royalty-free permission to reproduce all or portions of the above article/paper, and to authorize others to do so, for U.S. Government purposes.
• In the event the above article/paper is not accepted and published by AAAI, or is withdrawn by the author(s) before acceptance by AAAI, this agreement becomes null and void.

Information

• For Readers
• For Authors

Developed By

Part of the PKP Publishing Services Network

Copyright © 2021, Association for the Advancement of Artificial Intelligence. All rights reserved.

Information

• Author Services

Initiatives

You are accessing a machine-readable page. In order to be human-readable, please install an RSS reader.

All articles published by MDPI are made immediately available worldwide under an open access license. No special permission is required to reuse all or part of the article published by MDPI, including figures and tables. For articles published under an open access Creative Common CC BY license, any part of the article may be reused without permission provided that the original article is clearly cited. For more information, please refer to https://www.mdpi.com/openaccess .

Feature papers represent the most advanced research with significant potential for high impact in the field. A Feature Paper should be a substantial original Article that involves several techniques or approaches, provides an outlook for future research directions and describes possible research applications.

Feature papers are submitted upon individual invitation or recommendation by the scientific editors and must receive positive feedback from the reviewers.

Editor’s Choice articles are based on recommendations by the scientific editors of MDPI journals from around the world. Editors select a small number of articles recently published in the journal that they believe will be particularly interesting to readers, or important in the respective research area. The aim is to provide a snapshot of some of the most exciting work published in the various research areas of the journal.

Original Submission Date Received: .

• Active Journals
• Find a Journal
• Proceedings Series
• For Authors
• For Reviewers
• For Editors
• For Librarians
• For Publishers
• For Societies
• For Conference Organizers
• Open Access Policy
• Institutional Open Access Program
• Special Issues Guidelines
• Editorial Process
• Research and Publication Ethics
• Article Processing Charges
• Testimonials
• Preprints.org
• SciProfiles
• Encyclopedia

Article Menu

• Subscribe SciFeed
• Recommended Articles
• Google Scholar
• on Google Scholar
• Table of Contents

Find support for a specific problem in the support section of our website.

Please let us know what you think of our products and services.

Visit our dedicated information section to learn more about MDPI.

JSmol Viewer

A deep learning framework for evaluating the over-the-air performance of the antenna in mobile terminals.

1. Introduction

• The proposed framework introduces a novel antenna performance prediction model, which can be used for preliminary assessment of SAR and EIRP while reducing the number of repetitive simulations with commercial software on the specialized computational platform;
• The study converts the features of the antenna and the internal components of the mobile phone into a near-field EMF distribution on the Huygens’ box, which simplifies the input of the DL model;
• A DL model based on the divide-and-conquer method [ 15 ] is proposed. It divides the estimation for SAR and EIRP into eight modules, reducing the training complexity and enhancing the convergence rate;
• The Wilcoxon Signed Rank (WSR) test [ 16 ] for assessing SAR and TRP accuracy is developed. The Feature Selection Validation (FSV) [ 17 ] method is employed to evaluate EIRP accuracy. These methods provide a comprehensive understanding of the performance of the antenna.

2. The Proposed Method

2.1. the workflow of the proposed method.

• A phone in FS;
• A phone mounted in the “cheek” position on the head phantom (this includes the “cheek” position for both the left and right ears);
• PDA (Personal Digital Assistant) grip in “data” mode (for both the left and right hand);
• “Talk” mode (which includes configurations for the left and right head and hand).
• Additionally, four SAR evaluation configurations are created, including:
• The cheek (left and right) positions of the phone against the Specific Anthropomorphic Mannequin (SAM) phantom;
• Tilt (left and right) positions of the phone against the SAM phantom.

2.2. RTEEMF-PhoneAnts Architecture

3. experiment configurations, 3.1. dataset generation, 3.2. data preprocessing for rteemf-phoneants, 3.3. training configurations for rteemf-phoneants, 3.4. evaluation criteria and numerical validation, 4. experiment results, 4.1. performance of the proposed model, 4.2. efficiency of the proposed model, 4.3. model generalization across diverse antenna types, 5. conclusions, author contributions, institutional review board statement, informed consent statement, data availability statement, conflicts of interest, abbreviations.

• CTIA Certification. Test Plan for Wireless Device Over-the-Air Performance: Method of Measurement for Radiated RF Power and Receiver Performance ; CTIA Certification 3.8.2; CTIA Certification: Washington, DC, USA, 2019. [ Google Scholar ]
• IEC/IEEE 62704-1:2017 ; Determining the Peak Spatial-Average Specific Absorption Rate (SAR) in the Human Body from Wireless Communications Devices, 30 MHz to 6 GHz—Part 1: General Requirements for Using the Finite-Difference Time-Domain (FDTD) Method for SAR Calculations. IEC/IEEE International Standard: Piscataway, NJ, USA, 2017.
• IEC/IEEE 62704-2:2017 ; Determining the Peak Spatial-Average Specific Absorption Rate (SAR) in the Human Body from Wireless Communications Devices, 30 MHz to 6 GHz—Part 2: Specific Requirements for Finite Difference Time Domain (FDTD) Modelling of Exposure from Vehicle Mounted Antennas. IEEE/IEC International Standard: Piscataway, NJ, USA, 2017.
• IEC/IEEE 62209-1528:2020 ; Measurement Procedure for the Assessment of Specific Absorption Rate of Human Exposure to Radio Frequency Fields from Hand-Held and Body-Mounted Wireless Communication Devices—Part 1528: Human Models, Instrumentation, and Procedures (Frequency Range of 4 MHz to 10 GHz). IEC/IEEE International Standard: Piscataway, NJ, USA, 2020.
• Li, C.; Yang, L.; Wu, T. Predicting Over-the-Air Performance Using a Hand Phantom with Large Dielectric Flexibility. IEEE Trans. Electromagn. Compat. 2021 , 63 , 1699–1708. [ Google Scholar ] [ CrossRef ]
• Li, C.; Wu, T. Efficient Evaluation of Incident Power Density by Millimeter-Wave MIMO User Equipment Using Vectorized Field Superposition and Stochastic Population Optimizers. IEEE Trans. Electromagn. Compat. 2023 , 65 , 1090–1097. [ Google Scholar ] [ CrossRef ]
• Wu, T.; Chen, Y.; Li, C. Efficient Evaluation of Epithelial/Absorbed Power Density by Multi-Antenna User Equipment with SAM Head Model. IEEE Antennas Wirel. Propag. Lett. 2024 , 1–5. [ Google Scholar ] [ CrossRef ]
• Koziel, S.; Dabrowska, A.P. Performance-Based Nested Surrogate Modeling of Antenna Input Characteristics. IEEE Trans. Antenn. Propag. 2019 , 67 , 2904–2912. [ Google Scholar ] [ CrossRef ]
• Zhai, M.; Chen, Y.; Xu, L.; Yin, W.Y. An End-to-End Neural Network for Complex Electromagnetic Simulations. IEEE Antennas Wirel. Propag. Lett. 2023 , 22 , 2522–2526. [ Google Scholar ] [ CrossRef ]
• Khan, M.M.; Hossain, S.; Mozumda, P.; Akte, S.; Ashique, R.H. A review on machine learning and deep learning for various antenna design applications. Heliyon 2022 , 8 , e09317. [ Google Scholar ] [ CrossRef ]
• Love, A. The integration of the equations of propagation of electric waves. Phil. Trans. Roy. Soc. London Ser. A 1901 , 197 , 1–45. [ Google Scholar ]
• Scialacqua, L.; Mioc, F.; Scannavini, A.; Foged, L.J. Simulated and Measured Power Density using Equivalent Currents for 5G applications. In Proceedings of the 2020 IEEE International Symposium on Antennas and Propagation and North American Radio Science Meeting, Montreal, QC, Canada, 5–10 July 2020. [ Google Scholar ] [ CrossRef ]
• Eibert, T.F.; Kiliç, E.; Lopez, C.; Mauermayer, R.A.; Neitz, O.; Schnattinger, G. Electromagnetic Field Transformations for Measurements and Simulations (Invited Paper). Prog. Electromagn. Res. 2015 , 151 , 127–150. [ Google Scholar ] [ CrossRef ]
• Wang, H. Overview of Future Antenna Design for Mobile Terminals. Engineering 2022 , 11 , 12–14. [ Google Scholar ] [ CrossRef ]
• Li, C.; Yao, G.; Xu, X.; Yang, L.; Zhang, Y.; Wu, T.; Sun, J. DCSegNet: Deep Learning Framework Based on Divide-and-Conquer Method for Liver Segmentation. IEEE Access 2020 , 8 , 146838–146846. [ Google Scholar ] [ CrossRef ]
• Noether, G.E. Introduction to Wilcoxon (1945) individual comparisons by ranking methods. In Breakthroughs in Statistics: Methodology and Distribution ; Springer: New York, NY, USA, 1992; pp. 191–195. [ Google Scholar ]
• Duffy, A.P.; Martin, A.J.; Orlandi, A.; Antonini, G.; Benson, T.M.; Woolfson, M.S. Feature selective validation (FSV) for validation of computational electromagnetics (CEM). part I-the FSV method. IEEE Trans. Electromagn. Compat. 2006 , 48 , 449–459. [ Google Scholar ] [ CrossRef ]
• Liu, Z.; Hall, P.S.; Wake, D. Dual-frequency planar inverted-F antenna. IEEE Trans. Antenn. Propag. 1997 , 10 , 1451–1458. [ Google Scholar ]
• Khlouf, M.M.; Azemi, S.N.; Al-Hadi, A.A.; Soh, P.J.; Jamlos, M.F. Dual Band Planar Inverted F Antenna (PIFA) with L-Shape Configuration. MATEC Web Conf. 2017 , 140 , 01019. [ Google Scholar ] [ CrossRef ]
• Wang, H.; Zheng, M. A Triple-Band WLAN Monopole Antenna. In Proceedings of the Second European Conference on Antennas and Propagation, EuCAP 2007, Edinburgh, Scotland, UK, 11–16 November 2007. [ Google Scholar ]
• Wang, H.; Zheng, M.; Zhang, S.; Johnson, A. Monopole Slot Antenna. United States Patent US 661 8020B2 , 2003.
• Zheng, M.; Wang, H.; Hao, Y. Internal Hexa-Band Folded Monopole/Dipole/Loop Antenna with Four Resonances for Mobile Device. IEEE Trans. Antenn. Propag. 2012 , 60 , 2880–2885. [ Google Scholar ] [ CrossRef ]
• Rowell, C.; Lam, E.Y. Mobile-Phone Antenna Design. IEEE Antennas Propag. Mag. 2021 , 54 , 14–34. [ Google Scholar ] [ CrossRef ]
• Zhang, Z. Antenna Design for Mobile Devices ; Wiley–IEEE Press: Piscataway, NJ, USA, 2017. [ Google Scholar ]
• Chen, Z. Antennas for Portable Devices ; John Wiley & Sons, Inc.: New York, NY, USA, 2007. [ Google Scholar ]
• Jeong, S.; Cho, B.; Kwak, Y.S.; Byun, J.; Kim, A.S. Design and analysis of multi-band antenna for mobile handset applications. In Proceedings of the 2009 IEEE Antennas and Propagation Society International Symposium, North Charleston, SC, USA, 1–5 June 2009. [ Google Scholar ]
• Li, Y.; Zhang, Z.; Zheng, J.; Feng, Z. Compact Heptaband Reconfigurable Loop Antenna for Mobile Handset. IEEE Antennas Wirel. Propag. Lett. 2011 , 10 , 1162–1165. [ Google Scholar ] [ CrossRef ]
• Zuo, S.; Zhang, Z.; Xie, J. Design of dual-monopole slots antenna integrated with monopole strip for wireless wide area network mobile handset. IET Microw. Antenna Propag. 2014 , 8 , 194–199. [ Google Scholar ] [ CrossRef ]
• Singh, D.; Singh, B. Investigating the impact of data normalization on classification performance. Appl. Soft. Cmput. 2020 , 97 , 105524. [ Google Scholar ] [ CrossRef ]
• Duffy, A.; Zhan, G. FSV: State of the art and current research fronts. IEEE Electromagn. Compat. Mag. 2020 , 9 , 55–62. [ Google Scholar ] [ CrossRef ]
• Bongiorno, J.; Mariscotti, A. Uncertainty and Sensitivity of the Feature Selective Validation (FSV) Method. Electronics. 2022 , 11 , 2532. [ Google Scholar ] [ CrossRef ]
• 3rd Generation Partnership Project (3GPP). Evolved Universal Terrestrial Radio Access (E-UTRA). In User Equipment (UE) Radio Transmission and Reception ; 36.101 V18.5.0; 3GPP: Sophia Antipolis, France, 2024. [ Google Scholar ]
• Li, C.; Xu, C.; Wang, R.; Wu, T. Numerical evaluation of human exposure to 3. 5-GHz electromagnetic field by considering the 3GPP-like channel features. Ann. Telecommun. 2019 , 74 , 25–33. [ Google Scholar ] [ CrossRef ]
• Yuan, X. The Case for Electromagnetic Simulation in Mobile Antenna Design ; TECHNICAL PAPER; KYOCERA AVX Components Corporation: Fountain Inn, SC, USA, 2024. [ Google Scholar ]
• Cenci, M.P.; Eidelwein, E.M.; Scarazzato, T.; Veit, H.M. Assessment of smartphones’ components and materials for a recycling perspective: Tendencies of composition and target elements. J. Mater. 2024 , 26 , 1379–13931. [ Google Scholar ] [ CrossRef ]

Click here to enlarge figure

Share and Cite

Chen, Y.; Qi, D.; Yang, L.; Wu, T.; Li, C. A Deep Learning Framework for Evaluating the Over-the-Air Performance of the Antenna in Mobile Terminals. Sensors 2024 , 24 , 5646. https://doi.org/10.3390/s24175646

Chen Y, Qi D, Yang L, Wu T, Li C. A Deep Learning Framework for Evaluating the Over-the-Air Performance of the Antenna in Mobile Terminals. Sensors . 2024; 24(17):5646. https://doi.org/10.3390/s24175646

Chen, Yuming, Dianyuan Qi, Lei Yang, Tongning Wu, and Congsheng Li. 2024. "A Deep Learning Framework for Evaluating the Over-the-Air Performance of the Antenna in Mobile Terminals" Sensors 24, no. 17: 5646. https://doi.org/10.3390/s24175646

Article Metrics

Article access statistics, further information, mdpi initiatives, follow mdpi.

Subscribe to receive issue release notifications and newsletters from MDPI journals

arXiv's Accessibility Forum starts next month!

Help | Advanced Search

Computer Science > Multimedia

Title: deep learning-based video coding: a review and a case study.

Abstract: The past decade has witnessed great success of deep learning technology in many disciplines, especially in computer vision and image processing. However, deep learning-based video coding remains in its infancy. This paper reviews the representative works about using deep learning for image/video coding, which has been an actively developing research area since the year of 2015. We divide the related works into two categories: new coding schemes that are built primarily upon deep networks (deep schemes), and deep network-based coding tools (deep tools) that shall be used within traditional coding schemes or together with traditional coding tools. For deep schemes, pixel probability modeling and auto-encoder are the two approaches, that can be viewed as predictive coding scheme and transform coding scheme, respectively. For deep tools, there have been several proposed techniques using deep learning to perform intra-picture prediction, inter-picture prediction, cross-channel prediction, probability distribution prediction, transform, post- or in-loop filtering, down- and up-sampling, as well as encoding optimizations. In the hope of advocating the research of deep learning-based video coding, we present a case study of our developed prototype video codec, namely Deep Learning Video Coding (DLVC). DLVC features two deep tools that are both based on convolutional neural network (CNN), namely CNN-based in-loop filter (CNN-ILF) and CNN-based block adaptive resolution coding (CNN-BARC). Both tools help improve the compression efficiency by a significant margin. With the two deep tools as well as other non-deep coding tools, DLVC is able to achieve on average 39.6\% and 33.0\% bits saving than HEVC, under random-access and low-delay configurations, respectively. The source code of DLVC has been released for future researches.

Submission history

Access paper:.

• Other Formats

References & Citations

• Google Scholar
• Semantic Scholar

DBLP - CS Bibliography

Bibtex formatted citation.

Bibliographic and Citation Tools

Code, data and media associated with this article, recommenders and search tools.

• Institution

arXivLabs: experimental projects with community collaborators

arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.

Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.

Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs .

• DOI: 10.51878/cendekia.v4i2.2888
• Corpus ID: 270406076

ENGLISH SONGS FOR AN INDONESIAN TODDLER’ SECOND LANGUAGE ACQUISITION: A CASE STUDY

• Syifa Khofifah Saidah
• Published in CENDEKIA: Jurnal Ilmu… 11 June 2024
• CENDEKIA: Jurnal Ilmu Pengetahuan

Figures from this paper

7 References

Screen media and language development in infants and toddlers: an ecological perspective, incidental vocabulary learning through listening to songs, related papers.

Showing 1 through 3 of 0 Related Papers