by Bernard Marr
A more advanced implementation being explored would step in when a user makes a post saying they have, for example, a bicycle that they want to sell. It might automatically generate an advert-style post, discern the correct selling price based on the user's description and direct them to local selling pages where they may find a buyer.
This system is called Deep Text, because it relies on deep learning neural networks to analyze text and understand not just the words, but how the meaning of a word depends on its placement within a post and the other words used with it. This is a form of semi-unsupervised learning because rather than relying on a set of rules, such as a dictionary or a grammar rulebook, it learns for itself how words are used by “listening” to them – much the same way as a human does.
Suicide Prevention
Facebook also uses AI to monitor the way its users engage with the service, and looks for signs that individuals may be depressed or in danger of hurting themselves.8
It does this by looking for patterns in a user's posting behavior that match those of other posts that have previously been flagged as containing indicators that someone may be suicidal.
Signs could include users directly talking about themselves suffering or being unhappy, or receiving lots of messages from friends expressing concern or asking if they need help.
Once an alarm is raised, it is reviewed by human specialists before a decision is made on whether to intervene by offering the user information on how to receive help.
The social network doesn't currently contact users directly, preferring to put information at their fingertips in a timely way. But it has examined the possibility of alerting a user's real-world “support network” of friends and family. However, that would clearly have significant privacy implications.
FBLearner Flow
The “backbone” of Facebook's AI technology is its FBLearner Flow platform. It is designed to allow computer engineers to deploy AI in any area of the company's operation, without the engineer having to be a machine learning specialist.9
It is now in use by over 25% of Facebook's engineering teams, and is responsible for making 6 million predictions per second for the business and its customers. It is designed to be used to create algorithms that can easily be reused among multiple Facebook projects, once they have proven themselves effective.
Facebook AI Research
Facebook's machine learning research and development is coordinated through its Facebook AI Research division. Areas of research include ways in which smart, learning computer technology can be integrated with Facebook's services, how improvements can be made to core AI disciplines such as natural language processing and computer vision, and even how the future of socializing is likely to be shaped by augmented and virtual reality technology.
This year Facebook announced plans to grow the division to around 170 data scientists and engineers spread across its global offices, which include sites in Montreal, Pittsburgh, Paris, London and Tel Aviv.10
Key Challenges, Learning Points And Takeaways
The vast amount of information we share about our lives on Facebook means that the company has access to more of our personal data than just about anyone else.
Facebook has leveraged this to build features that keep us coming back to the site (and sharing more data), as well as match us with advertisers whose products we might want to buy.
All of this data – including our photos and text – has been invaluable to Facebook when it comes to training its facial recognition and natural language processing algorithms.
Unprecedented levels of insight into our lives means it can make increasingly accurate predictions about us – from what we want to buy to whether we are thinking about killing ourselves.
Notes
1Statistica, Number of monthly active Facebook users worldwide as of 2nd quarter 2018 (in millions): https://www.statista.com/statistics/ 264810/number-of-monthly-active-facebook-users-worldwide/
2Zephoria, Top 15 Valuable Facebook Statistics: https://zephoria.com/ top-15-valuable-facebook-statistics/
3Facebook, Introducing FBLearner Flow: Facebook's AI backbone: https://code.fb.com/core-data/introducing-fblearner-flow-facebook-s-ai-backbone/
4Facebook, Increasing Our Efforts to Fight False News: https://newsroom .fb.com/news/2018/06/increasing-our-efforts-to-fight-false-news/
5Facebook, Managing Your Identity on Facebook with Face Recognition Technology: https://newsroom.fb.com/news/2017/12/managing-your-identity-on-facebook-with-face-recognition-technology/
6Facebook, DeepFace: Closing the Gap to Human-Level Performance in Face Verification: https://research.fb.com/publications/deepface-closing-the-gap-to-human-level-performance-in-face-verification/
7Facebook, Introducing DeepText: Facebook's text understanding engine: https://code.fb.com/core-data/introducing-deeptext-facebook-s-text-understanding-engine/
8BBC, Facebook artificial intelligence spots suicidal users: https://www .bbc.co.uk/news/technology-39126027
9Facebook, Introducing FBLearner Flow: Facebook's AI backbone: https://code.fb.com/core-data/introducing-fblearner-flow-facebook-s-ai-backbone/
10Washington Post, Facebook, boosting artificial-intelligence research, says it's “not going fast enough”: https://www.washingtonpost.com/ technology/2018/07/17/facebook-boosting-artificial-intelligence -research-says-its-not-going-fast-enough/?utm_term=.de4f2c7f1298
7
IBM: Cognitive Computing Helps Machines Debate With Humans
IBM is the granddaddy of the computer industry, having been in existence for over 100 years. Constantly innovating, it dominated the mainframe industry in the 1960s and 1970s before pioneering the personal computer concept in the 1980s.
Like other US tech giants, it was not slow to understand the importance of machine learning. Its best known artificial intelligence (AI) endeavor is IBM Watson, a “cognitive computing” platform that became famous when it defeated two long-standing human champions at the gameshow Jeopardy!1
Since then Watson has been deployed across thousands of business use cases, and continues to be used by IBM to demonstrate the power and flexibility of its machine learning technology.
How Does IBM Use Artificial Intelligence?
As well as winning television gameshows, Watson has been deployed in many industries where its natural language processing capabilities are driving efficiency and creating new opportunities.
It was originally envisaged as a question-and-answer engine, but over the years its applications have diversified as its skillset has grown.
Royal Bank of Scotland uses Watson to power its customer service chatbot, Cora. Cora was trained on over 1,000 responses to 200 customer service queries. It then continues to learn after it is deployed, building links between natural language questions posed by human customers and the responses it has stored in its database.2 If a conversation becomes too difficult it will pass the customer to a human agent.
The rate at which Cora manages to handle queries by itself without having to resort to human assistance is known as the “containment rate” and is a key metric of its success. Currently, the rate is around 40% (and up to 80% for queries around commercial banking issues).3 The idea is that this rate will start to increase as the bot gains more experience of interacting with humans.
Stationery giant Staples used Watson to build a “smart ordering” system called the Easy Button. Essentially a voice-activated assistant similar to Amazon's Alexa, it is specifically trained in anticipating the stationery needs of Staples’ business customers. As it is repeatedly used it comes to learn about brands and quantities that customers require.4
Watson has broken into sports too. The All England Lawn Tennis Club worked with IBM at the world-famous Wimbledon tournaments to deliver automated highlights and enhance fan engagement. Trained with data from 22 years of tennis covering over 53 million data points, Watson was taught to deliver automated commentary, as well as real-time stats and analytics, directly to fans. A Wats
on-powered app called Ask Fred (named after Fred Perry) was also created to answer fans’ questions, from the history of tennis to where they can find public toilets at Wimbledon.5
Watson is also widely used in healthcare. The American Cancer Society used Watson to create the first AI assistant aimed at helping people diagnosed with cancer, and Watson For Oncology is a clinical support platform that advises doctors on treatment decisions, using thousands of pages of medical documents and case notes to predict treatment paths that are likely to have the best outcomes.6
If there was one task that common sense would seem to dictate that AI cannot yet do, it would be designing perfume. Global fragrance giant Symrise, which makes scents for Estee Lauder, Avon and Donna Karan, among others, thought differently. The result of their work with IBM is called Phylira, and it has developed scents – usually the preserve of human experts who have trained for years – that will soon be on sale in 4,000 Brazilian beauty stores.
Phylira works by breaking down scents into their constituent parts – the different oils, chemicals and natural extracts that are used to add specific flavors to each perfume, 1.7 million of them all together. It then read in sales data and customer service data to draw links showing which combinations of scents were likely to be appealing to different demographic groups.
The two fragrances that were developed by the algorithm achieved “stellar” results in focus group testing, proving more popular than other scents that had previously sold successfully to the target demographic (Brazilian millennials).7
Watson has certainly grown into a phenomenal success story for IBM since it stole the Jeopardy! crown. Aside from these use cases, Watson is also used by seven of the world's top 10 automotive companies and eight of the top 10 largest oil and gas companies.8
Project Debater
Perhaps the most impressive application of IBM's language processing AI technology is found in its Project Debater.
IBM says that Project Debater is the first AI system that can debate humans on complex subjects. It uses language processing and a database of hundreds of millions of articles covering 100 subject areas.9
It uses these tools and data to listen to its opponent's point of view, considers it, and then challenges it on logical and ethical grounds.
In its first live, public debate, Project Debater took on two experienced college debaters on the topics of whether space exploration should be government subsidized, and whether more telemedicine (medical practice carried out remotely by a doctor) would be a good thing.
On the subject of telemedicine, the audience voted that IBM's AI put forward a more compelling argument than its human opponent.10
Although overall the event was considered a draw, it marks a meaningful step forward for AI language processing. The technology has progressed from recognizing individual words, as has been done by email spam filtering for decades, to being able to answer basic questions (as seen in Siri and Alexa), to being able to engage in open, free-form debate.
Rather than analyzing a human sentence semantically, and trying to figure out what it wants to know, it must be able to discern a point that is being made, then construct an argument against it. It can do this either by citing authoritative facts that suggest the original point is based on bad information, or by finding logical flaws in the statements being made.
This technique of language processing is known as argument mining. In IBM's use case, it further broke this down into argument detection and argument stance classification. The former analyzes the argument to determine the claims that are being made and the evidence they are based on. The latter determines where those argument components sit in relation to the polarity of the discussion.11
It's worth noting that although Project Debater appears to be able to tackle any subject, it's still an example of specialized AI, rather than the fully generalized AI, which is still not likely to be around for some time. Although it has expert level knowledge on many subjects, it is only trained to apply that knowledge to debate. It would require further training for it to be able to use it for other purposes, for example, education.
While it mainly serves as an impressive display of AI competency right now, in the future IBM theorizes that the rules it is built on (and those it develops itself) will help humans choose decision outcomes that are evidence based, rather than influenced by bias, faulty logic or ambiguity.
Key Challenges, Learning Points And Takeaways
Thousands of businesses are using IBM Watson to take advantage of AI. Particularly active areas of work include customer relations, chatbots and medicine.
By focusing on language processing capabilities, IBM's strategy is to break down communication barriers between people and machines, making it easier for us to harness their potential.
IBM uses gameplay to demonstrate that its cognitive systems are capable of learning to solve puzzles in the same way that humans do, and with practice can become better than them. This began with Deep Blue's defeat of Kasparov and continues with Project Debater.
Project Debater represents AI evolving past its current ability to answer questions, and towards being able to engage in natural human conversations. This could have all sorts of implications for the future of AI.
Notes
1Tech Republic, IBM Watson: The inside story of how the Jeopardy-winning supercomputer was born, and what it wants to do next: https://www.techrepublic.com/article/ibm-watson-the-inside-story-of-how-the-jeopardy-winning-supercomputer-was-born-and-what-it-wants-to-do-next/
2IBM, Raising Cora: https://www.ibm.com/industries/banking-financial -markets/front-office/chatbots-banking
3IBM, Putting Smart to Work: https://www.ibm.com/blogs/insights-on-business/banking/putting-smart-work-raising-cora/
4IBM, How Staples is making customer service “easy” with Watson Conversation: https://www.ibm.com/blogs/watson/2017/02/staples-making -customer-service-easy-watson-conversation/
5IBM, How Wimbledon is using IBM Watson AI to power highlights, analytics and enriched fan experiences: https://www.ibm.com/blogs/ watson/2017/07/ibm-watsons-ai-is-powering-wimbledon-highlights- analytics-and-a-fan-experiences/
6American Cancer Society, American Cancer Society and IBM Collaborate to Create Virtual Cancer Health Advisor: http://pressroom. cancer.org/WatsonACSLaunch
7Vox, Is AI the future of perfume? IBM is betting on it: https://www. vox.com/the-goods/2018/10/24/18019918/ibm-artificial-intelligence- perfume-symrise-philyra
8IBM, IBM Largest Ever AI Toolset Release Is Tailor Made for 9 Industries and Professions: https://newsroom.ibm.com/2018-09-24-IBM-Largest- Ever-AI-Toolset-Release-Is-Tailor-Made-for-9-Industries-and- Professions
9The Verge, What it's like to watch an IBM AI successfully debate humans: https://www.theverge.com/2018/6/18/17477686/ibm-project-debater-ai
10The Guardian, Man 1, machine 1: landmark debate between AI and humans ends in draw: https://www.theguardian.com/technology/2018/ jun/18/artificial-intelligence-ibm-debate-project-debater
11IBM, Project Debater Datasets: https://www.research.ibm.com/haifa/ dept/vst/debating_data.shtml
8
JD.com: Automating Retail With Artificial Intelligence
JD.com is one of the largest online retailers in China and a company that prides itself on high tech and artificial intelligence (AI) enabled processes, which include a drone delivery system, autonomous delivery vehicles and robot-automated fulfilment centers.
We talked to a lot of businesses when putting this book together and although they all have different ideas about the future of AI, there is one thing most of them are at pains to agree on: AI isn't here to threaten human jobs and make us redundant, but to augment our own abilities.
JD.com's founder Liu Qiangdong (also known as Richard Liu) is the exception. In a World Retail Congress 2018 interview he said: “I hope my company would be a 100% automation company. I hope that someday there will be no human beings any more. 100% operated by the AI and the robots.”1
You
might suspect he is simply being more honest than most tech CEOs, who in reality would love to be able to do away with soft, squishy and demanding humans entirely (in their businesses, at least).
However, it gives a telling insight into JD.com's strategy for rolling out AI, which has focused on using robotics to physically automate as much of its retail operations as possible.
What Does JD.com Use Artificial Intelligence For?
JD.com's big push into AI has focused on deploying it to handle delivery, logistics and supply chain tasks across its vast retail network.
In fact, at its flagship Shanghai fulfilment center, which processes 200,000 orders per day, it employs a total of four people.2
Robots, powered by machine learning, move crates of products onto snaking networks of conveyor belts, which distribute items that are ready to be packed to other robots, which box them up and despatch them for delivery.
The extent to which AI has been built into JD.com's logistics is what has made it possible for them to offer a next-day delivery service to virtually any of China's 1.3 billion residents, no matter where in the country's 10 million km2 of territory they live. Now they are making preparations for the jump to same-day delivery.3
Of course, it is also using AI to improve customer experience. The company has even produced a chatbot that is capable of producing an automated piece of poetry, to be supplied to the recipient when items are purchased as gifts. The buyer can input characteristics of the person who will receive the gift and details of the occasion, and the robot will do the rest – how romantic!