by Bernard Marr
Even the best human radiologists are susceptible to fatigue and human error when their job involves examining hundreds of images every day. AI never gets tired, and providing it has accurate training data will not make mistakes or misdiagnoses.
Infervision founder and CEO Chen Kuan told me: “What I saw was that a lot of Chinese people, particularly those living outside big cities, do not get to have any regular medical check-ups involving medical imaging.
“So, they often have to wait until they feel something wrong with their body before they go to a big hospital where it can be diagnosed – by then it's often too late to do anything about it.
“So, what we wanted to do is use deep learning to alleviate this huge problem. If we can use it to learn from the past and assist in diagnosing more accurately, we can help solve the problem.”
Detecting cancer was the first application of Infervision's technology. Next, Kuan's team turned their attention to tackling another major killer – strokes.
What Technology, Tools And Data Were Used?
The primary source of data is medical image records, specifically patient X-ray and CT scans – Infervision has processed over 100,000 of each type of image.4
Infervision uses a supervised deep learning model, which means that it is trained on datasets where the outcome is known. In this case, it means the deep neural networks powering the algorithms are fed medical imagery from patients who had been given a cancer diagnosis. From images of healthy lungs, the system establishes a “normal” baseline – then is able to ring alarm bells when it comes across data that lies outside of the boundaries of what it considers to be normal.
As it processes more data (medical imagery) the algorithms “learn” to become more efficient at spotting cancer at an earlier stage, as they learn more and more about how signs of cancer present themselves.
The fact that the platform is able to use X-ray and CT scan images, rather than just MRI scans, is significant, because they are far cheaper to produce and are available to more people. MRI scans require more expensive machinery and more hours of work by specially trained humans, meaning they are less readily available, particularly in areas that aren't served by major hospitals.
What Were The Results?
The company has announced that it has partnerships with over 200 hospitals around the world and the technology is currently being used to analyze 20,000 scans every day.
Eliot Siegel, chairman of the Radiological Society of North America's Medical Imaging Resource Committee, said: “The application of AI will lead to a real digital shift in traditional medical imaging, requiring AI and people to work together to meet the challenges of the medical industry.
“In the process of lung nodule screening, Infervision is providing preemptive solutions that allow doctors to meet patients’ needs in a short period of time.”5
After proving the technology works for detecting lung cancers, Infervision has moved on to apply it to detecting early signs of strokes – another major killer. It is also working on detecting other types of cancer.
Key Challenges, Learning Points And Takeaways
Computer vision isn't just for fun image searches and marketing – it has the potential to save lives.
Deep neural networks enable computer algorithms to become increasingly efficient at sorting images according to how far they stray from a “normal” baseline.
New value can be extracted from old data when cutting-edge technologies such as deep learning are applied to existing datasets – much of the X-ray imagery used to train Infervision's platform was generated during the SARS outbreak of 2003.
Infervision stresses that its technology is not meant to replace doctors, but rather to enable them to work far more quickly and efficiently than they previously could.
Notes
1World Atlas, Leading causes of death in China: https://www.worldatlas. com/articles/leading-causes-of-death-in-china.html
2Forbes, How AI and Deep Learning is now used to Diagnose Cancer: https://www.forbes.com/sites/bernardmarr/2017/05/16/how-ai-and-deep-learning-is-now-used-to-diagnose-cancer/#24e50af6c783
3Infervision, About Us: http://www.infervision.com/Infer/aboutUS-en
4TechCrunch, Chinese startup Infervision emerges from stealth with an AI tool for diagnosing lung cancer: https://techcrunch.com/2017/05/08/ chinese-startup-infervision-emerges-from-stealth-with-an-ai-tool-for- diagnosing-lung-cancer/
5Digital Journal, Infervision Reaches 200-Hospital Milestone, Advances Global Medical Imaging Capabilities: http://www.digitaljournal.com/ pr/3928429
38
Mastercard: Using Artificial Intelligence To Cut Down The “False Declines” That Cost Businesses Billions Each Year
Mastercard processes billions of transactions every year, forming a crucial link between thousands of banks and millions of merchants.
In 2017, it acquired Brighterion to complete its mission of rolling out artificial intelligence (AI) technology across its entire network.
Its aim was to enable automated, machine learning-driven decision making at point of sale, online and offline, to ensure a secure yet smooth shopping experience at the point consumers hand over their card details.
What Problem Is Artificial Intelligence Helping To Solve?
“False declines” occur when a legitimate card transaction is incorrectly declined due to being flagged as suspicious and potentially fraudulent. As well as being inconvenient for customers, who in an increasingly cashless society may be left with no alternative way to pay, they cost US businesses $118 billion per year in lost revenue. This staggering figure is some 13 times higher than the actual cost of the fraud.1
These false positives have serious negative consequences for consumers’ brand loyalty – Mastercard's research found that one-third of us have stopped shopping at a retailer due to having a payment declined for apparently no good reason.2
Although false declines are expensive and inconvenient, using traditional methods of payment verification, based on static rules and datasets, they are somewhat inevitable.
How Is Artificial Intelligence Used In Practice?
Mastercard applies machine learning to the decision scoring system that is executed when a merchant's terminal passes a customer's card details through to Mastercard's issuers’ systems for verification at the point of purchase.3
This means that the models that are used to decide whether a payment is legitimate are updated in real time, based on data gathered from all of the billions of transactions that Mastercard processes. It works by building up a picture of how a card is used over time, allowing the algorithms to learn what is within normal boundaries of behavior and what could indicate suspicious activity.
Mastercard calls this system Decision Intelligence, and Ajay Bhalla, the company's president of enterprise risk and security, tells me that AI has created a real-time system that helped the business catch billions of dollars’ worth of fraud.
What Technology, Tools And Data Were Used?
Decisions are based mainly on logs of transactional data. This includes where a card is used, the volume of transactions carried out, the type of goods and services being purchased and information about the merchant where the card is being used.
It adds in what it knows about fraud trends and patterns around the world – when and where particular types of fraudulent transactions are likely to occur, as well as the types of business that are likely to be targeted.
It also uses anonymized and aggregated personal data relating to the person making the transaction.
Mastercard relies on open source AI solutions for some of its work, although the majority of the heavy lifting is done using its own proprietary algorithms developed in house and by Brighterion. It relies on both supervised (using labelled data) and unsupervised (unlabelled data) learning techniques for training of its algorithms.
What Were The Results?
Bhalla says that since the rollout of its network-wide AI platform, the organization
has increased its success rate at detecting fraud threefold, and the number of false positives has been reduced by roughly 50%.
Key Challenges, Learning Points And Takeaways
Decisioning based on static datasets using fixed rules is not sufficient for fast, hassle-free fraud verification over a network of Mastercard's scale.
Datasets and predictive models that update in real time allow for far more accurate predictions about the legitimacy of a transaction, meaning fewer false declines.
When Mastercard made the decision to implement AI across its entire network, talent acquisition proved to be a challenge. It overcame this by making acquisitions such as Brighterion, bringing their expertise on board.
Data quality is of utmost importance in initiatives like Mastercard's – inaccurate data would lead to a potentially even greater number of false positives, or fraudulent transactions being incorrectly approved. Both would reduce trust in the network and lead to financial losses. Because the system relies primarily on Mastercard's transactional data, it is considered highly reliable.
Notes
1Mastercard, MasterCard IQ Series Minimizes False Payment Declines: https://newsroom.mastercard.com/mea/press-releases/mastercard-iq-series-minimizes-false-payment-declines/
2Mastercard, Decision Intelligence: https://www.flickr.com/photos/ mastercardnews/31335572915/sizes/l
3Mastercard, Mastercard Rolls Out Artificial Intelligence Across its Global Network: https://newsroom.mastercard.com/press-releases/mastercard-rolls-out-artificial-intelligence-across-its-global-network/
39
Salesforce: How Artificial Intelligence Helps Businesses Understand Their Customers
Salesforce is one of the world's leading suppliers of customer relationship management (CRM) solutions. Its products and services are built to help businesses grow and track relationships with their customers.
When it was founded in 1999, Salesforce pioneered the concept of software-as-a-service (SaaS) delivered over the internet (now most commonly referred to as “in the cloud”).
Rather than selling or licensing software packages for customers to install and run autonomously on their own machines, SaaS providers charge a subscription fee for customers to access software running on their servers.
For vendors, this provides an ongoing revenue stream. For customers, it removes the hassle of having to maintain and update installations and deal with compatibility issues, while generally also lowering initial setup and deployment costs.1
With the emergence of cognitive, smart computing platforms as a dominant trend in business IT, it is now using the same cloud delivery model to arm its customers – businesses – with artificial intelligence (AI), and help them in turn understand and better manage their own customer bases.
What Problem Is Artificial Intelligence Helping To Solve?
Today's businesses face the challenge of maintaining customer relationships across a plethora of channels – from old fashioned mail shots to social media and chatbots, often trying to acquire and retain customers across different parts of the world. This increased complexity, involving multifaceted and fast-changing datasets, is ideal for AI.
The problem, though, is that building an AI infrastructure from the ground up – developing tools, training algorithms and gathering data – can be difficult and expensive.
How Is Artificial Intelligence Used In Practice?
Salesforce offers its business customer its Salesforce Einstein platform, which it calls the world's only comprehensive AI solution for CRM.2
Einstein is integrated into Salesforce's Customer Success Platform – the collective name for the various components of its cloud-hosted CRM solution.
Along with other AI-as-a-service providers it aims to put self-learning computing technology in the hands of businesses of just about any size. When Salesforce started working towards this goal, however, it realized that it faced one particularly tricky challenge. Businesses are understandably very protective of their customer data. Even with the promised advances that AI could bring, would they be able to persuade their customers to upload their data to a cloud that wasn't exclusively under their own control?
Crucially, Salesforce's data scientists and engineers were able to come up with a solution that meant they wouldn't have to. They designed their machine learning algorithms to work with metadata rather than the actual data itself.
This means that they automated the data preparation process, using machine learning to label elements of customers’ data – for example, by recognizing if a field in their CRM database contained an email or a marketing objective.
This effectively allows customers to run their data through Einstein's predictive machine learning algorithms without the algorithms themselves being able to “see” it.
However, if businesses are happy to allow other companies to benefit from the insights contained in their data, they can “opt out” of these features, meaning their (anonymized) data will be fed into algorithms powering services that use aggregated, pooled data.3
Einstein-powered services are available to help customers run many different business processes, including sales and marketing, invoicing and financial planning, community management and customer service.
What Technology, Tools And Data Were Used?
Einstein's training data is all of the information that businesses store relating to their customers, including transactional records and details of customer service interactions.
This also includes data gathered through its cloud enterprise services such as its collaborative working environment Chatter, and its email, calendar apps and social data streams.4
Salesforce built a team of 175 data scientists and spent $4 billion acquiring specialist AI businesses, including Metamind, RelateIQ and BeyondCore, as it developed its Einstein technology.5
Most recently, it has equipped Einstein with natural language processing technology, giving it the ability to understand its users’ voice commands. This means common tasks such as running analytical queries and reviewing CRM objectives can be done without touching a keyboard.6
What Were The Results?
With Einstein, Salesforce has effectively positioned itself as the first provider of AI-as-a-service for CRM.
Companies engaging with the service can expect to leverage the power of AI to acquire greater insights into marketing and customer service issues.
In the long term, this should lead to more satisfied customers providing higher lifetime value to the business.
Key Challenges, Learning Points And Takeaways
Delivering AI-as-a-service has the potential to drive strong economic growth by empowering businesses of any size to take advantage of these powerful tools and technologies.
Customer relations can effectively be managed in an automated way via machine learning by algorithms that can learn the most effective approaches to marketing and managing relationships with individuals according to the profile they fit.
Salesforce makes data ownership a unique selling point of their service, meaning that their customers do not have to let their valuable customer data out of their hands to take advantage of its cloud-based services.
Notes
1CIO, Software as a Service (SaaS) Definition and Solutions: https://www. cio.com/article/2439006/web-services/software-as-a-service–saas– definition-and-solutions.html
2Salesforce, FAQ: https://www.salesforce.com/uk/products/einstein/faq/
3Computer World, How Salesforce brought artificial intelligence and machine learning into its products with Einstein: https://www.computer- worlduk.com/cloud-computing/how-salesforce-brought-ai-machine- learning-into-its-platform-of-products-3647570/
4Computer World, What is Salesforce Einstein? Latest features & pricing: https://www.computerworlduk.com/cloud-computing/what-is-salesforce-einstein-3646520/
5Computer World, The biggest AI and machine learning acquisitions 2016: From Apple to Google, breaking down the AI acquisition binge: https://www.com
puterworlduk.com/galleries/it-business/biggest-ai-machine-learning-acquisitions-2016-3645450/
6Venturebeat, Salesforce announces Einstein Voice, a voice assistant for enterprises: https://venturebeat.com/2018/09/19/salesforce-announces-einstein-voice-a-voice-assistant-for-enterprises/
40
Uber: Using Artificial Intelligence To Do Everything
Uber built its business model around disruptive use of data – pairing public hire drivers with passengers by correlating location data from both parties’ smartphones. This meant it could assign drivers to waiting passengers far more quickly than the traditional taxi companies whose business it disrupted so dramatically.
Although it may not seem like it, Uber has been around for nearly 10 years now and in that time it has increasingly invested in artificial intelligence (AI).
In fact, it has been referred to as the first “AI-first” company,1 meaning that every business function from marketing to the core business function of providing passenger journeys is built on AI.
What Problems Is Artificial Intelligence Helping To Solve?
A major challenge faced by traditional taxi hire businesses is how to efficiently get customers home while incurring a minimum of expense through driver wages and mileage.