Ancenored

315

Deep Learning and Cognitive Computing

LEARNING OBJECTIVES

■■ Learn what deep learning is and how it is changing the world of computing

■■ Know the placement of deep learning within the broad family of artificial intelligence (AI) learning methods

■■ Understand how traditional “shallow” artificial neural networks (ANN) work

■■ Become familiar with the development and learning processes of ANN

■■ Develop an understanding of the methods to shed light into the ANN black box

■■ Know the underlying concept and methods for deep neural networks

■■ Become familiar with different types of deep learning methods

■■ Understand how convolutional neural networks (CNN) work

■■ Learn how recurrent neural networks (RNN) and long short-memory networks (LSTM) work

■■ Become familiar with the computer frameworks for implementing deep learning

■■ Know the foundational details about cognitive computing

■■ Learn how IBM Watson works and what types of application it can be used for

A rtificial intelligence (AI) is making a re-entrance into the world of commuting and in our lives, this time far stronger and much more promising than before. This unprecedented re-emergence and the new level of expectations can largely be attributed to deep learning and cognitive computing. These two latest buzzwords de- fine the leading edge of AI and machine learning today. Evolving out of the traditional artificial neural networks (ANN), deep learning is changing the very foundation of how machine learning works. Thanks to large collections of data and improved computational resources, deep learning is making a profound impact on how computers can discover complex patterns using the self-extracted features from the data (as opposed to a data scientist providing the feature vector to the learning algorithm). Cognitive computing— first popularized by IBM Watson and its success against the best human players in the game show Jeopardy!—makes it possible to deal with a new class of problems, the type

C H A P T E R

6

316 Part II • Predictive Analytics/Machine Learning

of problems that are thought to be solvable only by human ingenuity and creativity, ones that are characterized by ambiguity and uncertainty. This chapter covers the concepts, methods, and application of these two cutting-edge AI technology trends.

6.1 Opening Vignette: Fighting Fraud with Deep Learning and Artificial Intelligence 316

6.2 Introduction to Deep Learning 320 6.3 Basics of “Shallow” Neural Networks 325 6.4 Process of Developing Neural Network–Based Systems 334 6.5 Illuminating the Black Box of ANN 340 6.6 Deep Neural Networks 343 6.7 Convolutional Neural Networks 349 6.8 Recurrent Networks and Long Short-Term Memory Networks 360 6.9 Computer Frameworks for Implementation of Deep Learning 368

6.10 Cognitive Computing 370

6.1 OPENING VIGNETTE: Fighting Fraud with Deep Learning and Artificial Intelligence

THE BUSINESS PROBLEM

Danske Bank is a Nordic universal bank with strong local roots and bridges to the rest of the world. Founded in October 1871, Danske Bank has helped people and businesses in the Nordics realize their ambitions for over 145 years. Its headquarters is in Denmark, with core markets in Denmark, Finland, Norway, and Sweden.

Mitigating fraud is a top priority for banks. According to the Association of Certified Fraud Examiners, businesses lose more than $3.5 trillion each year to fraud. The problem is pervasive across the financial industry and is becoming more prevalent and sophis- ticated each month. As customers conduct more banking online across a wider variety of channels and devices, there are more opportunities for fraud to occur. Adding to the problem, fraudsters are becoming more creative and technologically savvy—they are also using advanced technologies such as machine learning—and new schemes to defraud banks are evolving rapidly.

Old methods for identifying fraud, such as using human-written rules engines, catch only a small percentage of fraud cases and produce a significantly high number of false positives. While false negatives end up costing money to the bank, chasing after a large number of false positives not only costs time and money but also blemishes customer trust and satisfaction. To improve probability predictions and identify a much higher per- centage of actual cases of fraud while reducing false alarms, banks need new forms of analytics. This includes using artificial intelligence.

Danske Bank, like other global banks, is seeing a seismic shift in customer interac- tions. In the past, most customers handled their transactions in a bank branch. Today, almost all interactions take place digitally through a mobile phone, tablet, ATM, or call center. This provides more “surface area” for fraud to occur. The bank needed to mod- ernize its fraud detection defenses. It struggled with a low 40 percent fraud detection rate and was managing up to 1,200 false positives per day—and 99.5 percent of all cases the bank was investigating were not fraud related. That large number of false alarms required a substantial investment of people, time, and money to investigate what turned out to be dead ends. Working with Think Big Analytics, a Teradata company, Danske Bank made a strategic decision to apply innovative analytic techniques, including AI, to better identify instances of fraud while reducing false positives.

Chapter 6 • Deep Learning and Cognitive Computing 317

THE SOLUTION: DEEP LEARNING ENHANCES FRAUD DETECTION

Danske Bank integrated deep learning with graphics processing unit (GPU) appliances that were also optimized for deep learning. The new software system helps the analyt- ics team to identify potential cases of fraud while intelligently avoiding false positives. Operational decisions are shifted from users to AI systems. However, human interven- tion is still necessary in some cases. For example, the model can identify anomalies, such as debit card purchases taking place around the world, but analysts are needed to determine whether that is fraud or a bank customer simply made an online purchase that sent a payment to China and then bought an item the next day from a retailer based in London.

Danske Bank’s analytic approach employs a “champion/challenger” methodology. With this approach, deep learning systems compare models in real time to determine which one is most effective. Each challenger processes data in real time, learning as it goes which traits are more likely to indicate fraud. If a process dips below a certain threshold, the model is fed more data, such as the geolocation of customers or recent ATM transactions. When a challenger outperforms other challengers, it transforms into a champion, giving the other models a roadmap to successful fraud detection.

THE RESULTS

Danske Bank implemented a modern enterprise analytic solution leveraging AI and deep learning, and it has paid big dividends. The bank was able to:

• Realize a 60 percent reduction in false positives with an expectation to reach as high as 80 percent.

• Increase true positives by 50 percent. • Focus resources on actual cases of fraud.

The following graph (see Figure 6.1) shows how true and false positive rates improved with advanced analytics (including deep learning). The red dot represents the old rules engine, which caught only about 40 percent of all fraud. Deep learning improved signifi- cantly upon machine learning, allowing Danske Bank to better detect fraud with much lower false positives.

Enterprise analytics is rapidly evolving and moving into new learning systems enabled by AI. At the same time, hardware and processors are becoming more powerful and spe- cialized, and algorithms more accessible, including those available through open source. This gives banks the powerful solutions needed to identify and mitigate fraud. As Danske Bank learned, building and deploying an enterprise-grade analytics solution that meets its specific needs and leverages its data sources deliver more value than traditional off- the-shelf tools could have provided. With AI and deep learning, Danske Bank now has the ability to better uncover fraud without being burdened by an unacceptable amount of false positives. The solution also allows the bank’s engineers, data scientists, lines of business, and investigative officers from Interpol, local police, and other agencies to col- laborate to uncover fraud, including sophisticated fraud rings. With its enhanced capabili- ties, the enterprise analytic solution is now being used across other business areas of the bank to deliver additional value.

Because these technologies are still evolving, implementing deep learning and AI solutions can be difficult for companies to achieve on their own. They can benefit by partnering with a company that has the proven capabilities to implement technology- enabled solutions that deliver high-value outcomes. As shown in this case, Think Big Analytics, a Teradata company, has the expertise to configure specialized hardware and software frameworks to enable new operational processes. The project entailed integrat- ing open-source solutions, deploying production models, and then applying deep learning

318 Part II • Predictive Analytics/Machine Learning

analytics to extend and improve the models. A framework was created to manage and track the models in the production system and to make sure the models could be trusted. These models enabled the underlying system to make autonomous decisions in real time that aligned with the bank’s procedural, security, and high-availability guidelines. The solution provided new levels of detail, such as time series and sequences of events, to better assist the bank with its fraud investigations. The entire solution was implemented very quickly—from kickoff to live in only five months. Figure 6.2 shows a generalized framework for AI and deep learning–based enterprise-level analytics solutions.

In summary, Danske Bank undertook a multi-step project to productionize machine- learning techniques while developing deep learning models to test those techniques. The integrated models helped identify the growing problem of fraud. For a visual summary, watch the video (https://www.teradata.com/Resources/Videos/Danske-Bank- Innovating-in-Artificial-Intelligence) and/or read the blog (http://blogs.teradata. com/customers/danske-bank-innovating-artificial-intelligence-deep-learning- detect-sophisticated-fraud/).

Deep Learning

21.0

20.8

20.6

20.4

20.2

0.0

0.0 0.02

Tr ue

P os

it iv

e R

at e

0.04 0.06

False Negative Rate

Random predi ction

0.08 0.10

Classic Machine Learning

Rules Engine

Ensemble (area = 0.89) CNN (area = 0.95) ResNet (area = 0.94) LSTM (area = 0.90) Rule Engine Random predictions

FIGURE 6.1 Deep Learning Improves Both True Positives and True Negatives.

https://www.teradata.com/Resources/Videos/Danske-Bank-Innovating-in-Artificial-Intelligence

https://www.teradata.com/Resources/Videos/Danske-Bank-Innovating-in-Artificial-Intelligence

http://blogs.teradata.com/customers/danske-bank-innovating-artificial-intelligence-deep-learning-detect-sophisticated-fraud/

http://blogs.teradata.com/customers/danske-bank-innovating-artificial-intelligence-deep-learning-detect-sophisticated-fraud/

http://blogs.teradata.com/customers/danske-bank-innovating-artificial-intelligence-deep-learning-detect-sophisticated-fraud/

Chapter 6 • Deep Learning and Cognitive Computing 319

u QUESTIONS FOR THE OPENING VIGNETTE

1. What is fraud in banking? 2. What are the types of fraud that banking firms are facing today? 3. What do you think are the implications of fraud on banks and on their customers? 4. Compare the old and new methods for identifying and mitigating fraud. 5. Why do you think deep learning methods provided better prediction accuracy? 6. Discuss the trade-off between false positive and false negative (type 1 and type 2

errors) within the context of predicting fraudulent activities.

WHAT WE CAN LEARN FROM THIS VIGNETTE

As you will see in this chapter, AI in general and the methods of machine learning in specific are evolving and advancing rapidly. The use of large digitized data sources, both from inside and outside the organization, both structured and unstructured, along with advanced computing systems (software and hardware combinations), has paved the way toward dealing with problems that were thought to be unsolvable just a few years ago. Deep learning and cognitive computing (as the ramifications of the cutting edge in AI systems) are helping enterprises to make accurate and timely decisions by harnessing the rapidly expanding Big Data resources. As shown in this opening vignette, this new generation of AI systems is capable of solving problems much bet- ter than their older counterparts. In the domain of fraud detection, traditional methods have always been marginally useful, having higher than desired false positive rates and causing unnecessary investigations and thereby dissatisfaction for their customers. As difficult problems such as fraud detection are, new AI technologies like deep learn- ing are making them solvable with a high level of accuracy and applicability.

Source: Teradata Case Study. “Danske Bank Fights Fraud with Deep Learning and AI.” https://www.teradata. com/Resources/Case-Studies/Danske-Bank-Fight-Fraud-With-Deep-Learning-and-AI (accessed August 2018). Used with permission.

Engineer

Simulate

M entoring

Handover Investigate

Cross-Functional Teams

4

3 2

1

Cross-Functional Teams

Leveragable

APIs

Validate

InsightsLive Test

Production

Test

Integrate

Analyze Data

Go Live

Tr ain

ing

Ini tia

l W in

s

Al as-a-Service Manage iterative, stage-gate process for analytic models from development to handover to operations

Al Strategy Analyze business priorities and identify Al use cases. Review key enterprise AI capabilities and provide recommendations and next steps for customers to successfully get value from AI.

Al Rapid Analytic Consulting EngagementTM (Race) Use AI exploration to test use cases and provide a proof of value for AI approaches.

Al Foundation Operationalize use cases through data science and engineering; build and deploy a deep learning platform, integrating data sources, models, and business processes.

FIGURE 6.2 A Generalized Framework for AI and Deep Learning–Based Analytics Solutions.

https://www.teradata.com/Resources/Case-Studies/Danske-Bank-Fight-Fraud-With-Deep-Learning-and-AI

https://www.teradata.com/Resources/Case-Studies/Danske-Bank-Fight-Fraud-With-Deep-Learning-and-AI

320 Part II • Predictive Analytics/Machine Learning

6.2 INTRODUCTION TO DEEP LEARNING

About a decade ago, conversing with an electronic device (in human language, intelligently) would have been unconceivable, something that could only be seen in SciFi movies. Today, however, thanks to the advances in AI methods and technologies, almost everyone has ex- perienced this unthinkable phenomenon. You probably have already asked Siri or Google Assistant several times to dial a number from your phone address book or to find an address and give you the specific directions while you were driving. Sometimes when you were bored in the afternoon, you may have asked the Google Home or Amazon’s Alexa to play some music in your favorite genre on the device or your TV. You might have been surprised at times when you uploaded a group photo of your friends on Facebook and observed its tagging suggestions where the name tags often exactly match your friends’ faces in the pic- ture. Translating a manuscript from a foreign language does not require hours of struggling with a dictionary; it is as easy as taking a picture of that manuscript in the Google Translate mobile app and giving it a fraction of a second. These are only a few of the many, ever- increasing applications of deep learning that have promised to make life easier for people.

Deep learning, as the newest and perhaps at this moment the most popular member of the AI and machine-learning family, has a goal similar to those of the other machine- learning methods that came before it: mimic the thought process of humans—using math- ematical algorithms to learn from data pretty much the same way that humans learn. So, what is really different (and advanced) in deep learning? Here is the most commonly pronounced differentiating characteristic of deep learning over traditional machine learn- ing. The performance of traditional machine-learning algorithms such as decision trees, support vector machines, logistic regression, and neural networks relies heavily on the representation of the data. That is, only if we (analytics professionals or data scientists) provide those traditional machine-learning algorithms with relevant and sufficient pieces of information (a.k.a. features) in proper format are they able to “learn” the patterns and thereby perform their prediction (classification or estimation), clustering, or association tasks with an acceptable level of accuracy. In other words, these algorithms need humans to manually identify and derive features that are theoretically and/or logically relevant to the objectives of the problem on hand and feed these features into the algorithm in a proper format. For example, in order to use a decision tree to predict whether a given customer will return (or churn), the marketing manager needs to provide the algorithm with information such as the customer’s socioeconomic characteristics—income, occupa- tion, educational level, and so on (along with demographic and historical interactions/ transactions with the company). But the algorithm itself is not able to define such socio- economic characteristics and extract such features, for instance, from survey forms com- pleted by the customer or obtained from social media.

While such a structured, human-mediated machine-learning approach has been working fine for rather abstract and formal tasks, it is extremely challenging to have the approach work for some informal, yet seemingly easy (to humans), tasks such as face identification or speech recognition since such tasks require a great deal of knowledge about the world (Goodfellow et al., 2016). It is not straightforward, for instance, to train a machine-learning algorithm to accurately recognize the real meaning of a sentence spo- ken by a person just by manually providing it with a number of grammatical or semantic features. Accomplishing such a task requires a “deep” knowledge about the world that is not easy to formalize and explicitly present. What deep learning has added to the classic machine-learning methods is in fact the ability to automatically acquire the knowledge required to accomplish such informal tasks and consequently extract some advanced fea- tures that contribute to the superior system performance.

To develop an intimate understanding of deep learning, one should learn where it fits in the big picture of all other AI family of methods. A simple hierarchical relationship diagram,

Chapter 6 • Deep Learning and Cognitive Computing 321

or a taxonomy-like representation, may in fact provide such a holistic understanding. In an attempt to do this, Goodfellow and his colleagues (2016) categorized deep learning as part of the representation learning family of methods. Representation learning techniques entail one type of machine learning (which is also a part of AI) in which the emphasis is on learn- ing and discovering features by the system in addition to discovering the mapping from those features to the output/target. Figure 6.3 uses a Venn diagram to illustrate the place- ment of deep learning within the overarching family of AI-based learning methods.

Figure 6.4 highlights the differences in the steps/tasks that need to be performed when building a typical deep learning model versus the steps/tasks performed when building models with classic machine-learning algorithms. As shown in the top two work- flows, knowledge-based systems and classic machine-learning methods require data sci- entists to manually create the features (i.e., the representation) to achieve the desired output. The bottommost workflows show that deep learning enables the computer to derive some complex features from simple concepts that would be very effort intensive (or perhaps impossible in some problem situations) to be discovered by humans manu- ally, and then it maps those advanced features to the desired output.

From a methodological viewpoint, although deep learning is generally believed to be a new area in machine learning, its initial idea goes back to the late 1980s, just a few decades after the emergence of artificial neural networks when LeCun and colleagues (1989) published an article about applying backpropagation networks for recognizing handwritten ZIP codes. In fact, as it is being practiced today, deep learning seems to be nothing but an extension of neural networks with the idea that deep learning is able to deal with more complicated tasks with a higher level of sophistication by employing many layers of connected neurons along with much larger data sets to automatically character- ized variables and solve the problems but only at the expense of a great deal of compu- tational effort. This very high computational requirement and the need for very large data sets were the two main reasons why the initial idea had to wait more than two decades until some advanced computational and technological infrastructure emerged for deep