TEST 3: Reading Notes Ch. 15.7

supervised learning def

where algorithms are trained by providing explicit examples of results sought, like defective vs. error-free, or stock price

There are also many subcategories of machine learning and other types of AI training, including

supervised learning unsupervised learning semi-supervised learning

Understanding Popular Types of AI

...

Neural networks def

A statistical techniques used in AI, and particularly in machine learning. Neural networks hunt down and expose patterns, building multilayered relationships that humans can't detect on their own. Many refer to the multilayered interconnections among data as mimicking the neurons of the brain (hence the name). If a set of interrelationships is strong, they go into the pattern-matching scheme. If a better set of relationships is found, old ones are tweaked or discarded. Neural networks are often referred to as a "black box," meaning that the weights and relationships of data that identify patterns approximate a mathematical function, but are difficult to break out as you would in a traditional mathematical formula. Massive amounts of data play a role here. Google leverages data it collects when users talk to and type into its search engine to improve its speech recognition algorithms, cutting errors by 25 percent in a single rollout.

It's Not as Easy as the Press Might State: Technical, Organizational, Legal, and Societal Challenges of AI and Machine Learning

AI starts with what are sometimes referred to as "naked algorithms" that need to be trained. These are starting point algorithms, and many are in the public domain or accessible from cloud provider services (e.g., Google TensorFlow) or through vendor APIs and inside software development kits (e.g., Apple Core ML). However, training implies access to a large quantity of consistent, reliable historical data. Since most data mining is a subset of AI, many of the challenges mentioned in the prior section also apply to machine learning.

Expert systems def

AI systems that leverage rules or examples to perform a task in a way that mimics applied human expertise Expert systems are used in tasks ranging from medical diagnoses to product configuration. They may be programmed with explicit rules (think a big "if this, then do that" decision tree), or rules may be automatically built by analyzing specific cases against outcomes (e.g., make less product if the weather is below 40 degrees and rainy, since there will be less foot traffic).

CAPTCHAs def

An acronym standing for completely automated public Turing test to tell computers and humans apart. The Turing Test is, rather redundantly, an idea (rather than an official test) that one can create a test to tell computers apart from humans. (for completely automated public Turing test to tell computers and humans apart) are meant to keep out automated software that may create accounts used in spamming or other nefarious activity Google's widely-used method, known as reCAPTCHA, has asked users to prove their superior-to-robot chops by classifying images, such as hard-to-read letters, house numbers, and to identify squares containing stop lights and traffic signs. If you've taken part, you were contributing to massive databases that are used to help software "learn" by recognizing human-classified images. This has improved OCR

Key Takeaways

Artificial intelligence refers to software that can mimic or improve upon functions that would otherwise require human intelligence. Machine learning is a type of AI often broadly defined as software with the ability to learn or improve without being explicitly programmed. Deep learning is a subcategory of machine learning. The "deep" in deep learning refers to the layers of interconnections and analysis that are examined to arrive at results. Neural networks are a category of AI used to identify patterns by testing multilayered relationships that humans can't detect on their own. Neural network techniques are often used in machine learning and are popular in data mining. Expert systems are AI systems that leverage rules or examples to perform a task in a way that mimics applied human expertise. Genetic algorithms are model-building techniques where computers examine many potential solutions to a problem, iteratively modifying (mutating) various mathematical models, and comparing the mutated models to search for a best alternative function. The explosion of machine learning is due to several factors, including open source tools; cloud computing for low-cost, high-volume data crunching; tools accessible via API or as part of software development kits; and new chips specifically designed for the simultaneous pattern hunting and relationship testing used in neural networks and other types of AI. AI can now be found in all sorts of consumer products—it is enabling the first generation of autonomous vehicles; it's influencing soft-skill disciplines like human resources; it improves the efficiency of modern, complex operations; it's is a key component in many customer service initiatives; and it's even being used in medical diagnosis and research. Incorporating AI and machine learning remains challenging. Organizations require technical skills and must reeducate the workforce impacted by AI decision-making tools. Firms need a clean, deep, and accurate dataset. Predictions require a tight relationship between historical data and new data used for prediction. Legal issues may prohibit the use of machine learning and "black box" algorithms in industries requiring exposure of the decision-making process. Varying international laws may allow techniques in one region of the globe that are prohibited in others. AI and machine learning have exposed the possibility of unintended consequences, including the possibility of introducing unintended bias, discrimination, and violations of privacy.

AI is giving us better customer service and delivering many from "phone tree hell."

China Merchants Bank, uses an AI-based messaging bot in WeChat, China's most popular mobile platform, to handle 1.5 million to 2 million queries a day, a volume that would otherwise require a staff of 7,000 humans. Casino giant Caesars uses its Ivy virtual concierge to answer questions received via text, reducing calls to the human concierge desk by 30%. Airliner KLM has used customer service text bots to double the number of weekly customer inquiries to 120,000 while increasing the number of agents by only 6%. Insurers Humana and MetLife are using voice analysis software from the firm Cogito to identify if customer service reps have experienced "compassion fatigue." Identify an agent who shows signs an interaction may not be going well (speech speed, keywords, voice level), and the system can prompt service people with strategies to improve. This sort of advising-when-needed is especially useful in phone support, where employee turnover can easily exceed 40 percent a year. British grocer Ocado uses AI to scan 10,000 customer e-mails a day to escalate and prioritize response as well as identify sentiment (good vs. bad experiences) around keywords. Applications like this don't get rid of humans, but they make existing staff more efficient and productive.

artificial intelligence def

Computer software that can mimic or improve upon functions that would otherwise require human intelligence. The goal of AI = mimic or improve An explosion of tools is fueling the current spread of AI. These include a new generation of hardware chips that fuel AI through designs tailored to find patterns faster, cloud resources that any developer with a credit card can tap into, open source algorithms that can be applied to creating custom insights, software development kits that create standards for building AI into apps and other products, and data-capture tools that include sensors, cameras, and microphones. Google CEO Sundar Pichai says that AI will have a "more profound" impact than electricity or fire

AI might also help make the workplace fairer, analyzing employee promotion and pay data for anomalies that suggest managers have hired staff that looks "more like them" rather than those who exhibit the characteristics and track record for success.

For a look at AI for better hiring decisions, consider systems from the startup Pymetrics, used by firms as wide ranging as Unilever in packaged goods and Nielsen in ratings research. Pymetrics tests candidates on roughly 80 traits, including memory and attitude to risk, using machine learning to measure applicant scores against a firm's top performers and predict suitability and success.

Own an Apple Product? You're Already Using a Whole Lot of AI

Data collected and constantly analyzed by Siri helps the voice assistant continually improve with better understanding of voices and accents worldwide, develop a greater understanding of context, and parse how devices are used so it can better service requests. Use ApplePay? AI fights fraud by improving analysis with each transaction. Maps? AI helps plot the best route by analyzing all sorts of traffic input. HealthKit? Machine learning can help keep cheats from climbing the leaderboard by identifying legitimate activity and filtering out bogus movement. Like your photos? AI recognizes faces and builds photo collages on what are determined to be your "best" pictures. Used FaceID or chatted with friends as a talking poop animoji? AI identifies the contours of your face to know you are you, and how to turn you into a cartoon. AI is behind extending device battery life, auto-switching between cellular and Wi-Fi networks, choosing news stories, apps, music, and video content you might enjoy. The iPad relies on AI to tell the difference between movements from the Apple Pencil and accidental swipes and taps from the pencil holder's hand. And the HomePod speaker combines sensor data with AI to optimize sound to suit the acoustics of the room Expect even more since third parties have access to many of these tools, too. Apple's custom processors and the firm's Core ML software developer framework allow coders to tap into Apple and third-party machine learning algorithms so that apps can take advantage of image recognition, natural language processing, computer vision, and more. Tools like these create standards and prevent developers from having to create things from scratch, enabling even small-time programmers to cheaply, easily, and quickly incorporate AI in their products.

Issues that managers, as well as concerned citizens, might want to be aware of include:

Data quality, inconsistent data, or the inability to integrate data sources into a single dataset capable of input into machine learning systems can all stifle efforts. Not enough data. - A firm might want to get into machine learning, but may lack underlying databases to begin this effort. It might, for example, be impossible to use machine learning to develop a system to predict failure when there are very few cases of failure that occur, and hence not enough examples to learn from. This also applies to the inability to predict rare "black swan" events since, by definition, they are either exceedingly rare or have never previously occurred. Technical staff may require training in developing and maintaining such systems, and such skills are rare. - An IDC study reported a 50 percent failure rate in a fourth of over 2,700 firms surveyed in a sweeping AI study. In situations where AI makes a recommendation, but a human makes the final call, managers using such systems may need coaching on when to accept and when to question results (see the Tesco "milk loaf" example in the prior section). Machine learning may mean more human learning at all levels of the organization. AI systems also involve a discipline known as "change management" that goes hand-in-hand with many IS projects. Change management seeks to identify how workflows and processes are to be altered, and how to manage the worker and organizational transition from one system to another. This can be key because many users of corporate AI will see their jobs significantly altered. They might have to do more, take on more responsibility, or remove instinct from some decisions and rely on recommendations made by a machine. Some types of machine learning may be legally prohibited because of the data used or the inability to identify how a model works and whether or not it might be discriminatory. - For example, while gender and religion could be used to predict some risks, they are unacceptable to regulators in some applications and jurisdictions. Redlining laws in the lending industry prevent geography from being used in calculating credit worthiness, since geography is often tightly correlated with race. In other industries, regulators won't accept the "black box" solutions offered by neural networks. And some areas such as the EU may have higher privacy protection that prohibits the gathering or use of certain data or techniques. The negative unintended consequences of data misuse might also lead to regulation that limits techniques currently used. - Some believe this helps give China an edge in some systems, since the government keeps a vast database of faces that can help train facial-recognition algorithms, and privacy is less of a concern than in the West. Jaywalkers in Shanghai can already be fined (or shamed) from facial recognition that identifies scofflaw citizens. In another example, The Chinese financial firm Ping An uses app-based video interviews to spot shifty behavior worthy of further screening. Prospective borrowers answer a series of questions related to income and ability to replay a loan, while machine learning systems monitor and identify some fifty distinct facial expressions related to truthfulness. The camera and the cloud are becoming a sort of real-time lie detector. Many workers are startled to find that in the United States, just about anything done on organizational networks or using a firm's computer hardware can be monitored. - While examining worker communications can help ensure employees don't break the law or commit crimes against the firm, and can offer help on how to do one's job better, the acceleration of these practices will undoubtedly raise additional privacy issues and have the potential to alienate workers, especially in a tight labor market. And as we think of how data relates to competitive advantage, firms that gain an early lead and benefit from scale may be in a position to collect more data than competitors, fueling a virtuous cycle where early winners generate more data, have stronger predictive capabilities, and can have an edge in entering new markets, offering new services, attracting customers, and cutting prices. Good for the winners and possibly good for consumers in the short run, but this may also fuel the kind of winner-take-all / winner-take-most dominance we see when network effects are present, something that might stifle innovation if it discourages competition and feeds near-monopolies. Indeed, many have referred to data as "the new oil," in that it is has the ability to create cash-gushing opportunities.

Artificial Intelligence, Machine Learning, and Deep Learning

Deep learning is a subset of machine learning, which is a subset of artificial intelligence. These techniques involve some form of pattern recognition and are used in many applications

While AI is not a single technology—terms and categorizations may overlap or have debated definitions—various forms of AI can show up as part of analytics products, CRM tools, transaction processing systems, and other information systems. M

Many of these techniques that crunch massive amounts of data leverage the special AI-powering graphics chips and FPGAs mentioned in the Moore's Law and More chapter.

OCR def

Optical Character Recognition. Software that can scan images and identify text within them. t's improved the accuracy of Google Street View in recognizing house numbers, and it's enhancing the computer-vision algorithms used in driverless cars

AI is increasingly embedding itself into the soft-skills discipline of human resources.

Software from the SaaS vendor Workday provides software to improve employee retention by uncovering patterns of those most likely to leave. The firm's software examines some 60 factors, including salary, time between holidays taken, and turnover of the employee's manager. Another firm, Arena, provides technology to hospitals and home-care companies to use application and third-party data to screen job applicants for traits that indicate they may stay at their job longer. The firm claims median turnover has been reduced by nearly 40 percent in firms using their software.

Machine learning def

Subset of AI Results improved without explicit programming A type of artificial intelligence that leverages massive amounts of data so that computers can improve the accuracy of actions and predictions on their own without additional programming. a type of AI often broadly defined as software with the ability to learn or improve without being explicitly programmed Many of the data mining techniques described in the prior section use machine learning.

There's also a good chance you've been reading news items and other copy written by robo-journalists.

The Associated Press uses software called Wordsmith by the firm Automated Insights to craft more than 3,000 financial reports per quarter, posting summaries online within minutes of their release. The Los Angeles Times uses "Quakebot" to analyze geological data, in one case breaking the story of a 4.7 magnitude earthquake in Southern California. The Big Ten Network uses software from Narrative Science for updates of football and basketball games and for short recaps of collegiate baseball and softball. Box scores and other play-by-play data are all the inputs these systems need.

Alphabet reported its DeepMind AI subsidiary lost half a billion dollars in 2019, despite doubling sales.

The cost to acquire top tech talent is one reason cited for the loss. Google and Facebook have hired top university faculty with salary increases of 10 times more than they were paid in academic jobs. Mergers and acquisitions activity related to AI is also red hot. Part of the acquisition frenzy in AI is a so-called aqui-hire play to bring in hard-to-lure talent. Startups without revenue are being priced at $5 million to $10 million for each AI expert that comes along with the deal. Yet despite interest, firms are struggling to capitalize on trends and technologies. Some 85 percent of companies think AI will offer a competitive advantage, but only one in 20 is extensively employing it today.

A system developed by IBM and the Baylor College of Medicine "read" 100,000 research papers in two hours and "found completely new biology hidden in the data"—undiscovered, needle-like insights buried in the existing massive haystacks of human-generated knowledge, which no single researcher could digest and map.

These specific insights "could provide routes to new cancer drugs." Some hear about machine learning and have nightmares of Skynet from the Terminator movies (Elon Musk, Bill Gates, and Stephen Hawking are among well-known tech-forward crowds who have cautioned about AI's future), but big brain insights from AI just might save your life one day.

Deep learning def

Type of machine learning Includes several layers of analysis between input data and output results A type of machine learning that uses multiple layers of interconnections among data to identify patterns and improve predicted results. Deep learning most often uses a set of techniques known as neural networks and is popularly applied in tasks like speech recognition, image recognition, and computer vision. The "deep" in deep learning typically refers to the layers of interconnections and analysis that are examined to arrive at results. Some technologists refer to "deep" as having more than one "hidden" layer between input and output. That might not mean much to most managers, but just as "Big Data" is "a lot," "Deep Learning" has more analytical complexity.

Learning Objectives

Understand terms such as artificial intelligence, machine learning, and deep learning. Name, recognize, and give examples of popular types of AI used in modern organizations. Understand the rise of AI and machine learning, as well as the challenges associated with leveraging this technology in an organization, and its broader social implications.

Examples of AI in Action

While the current uses of artificial intelligence and machine learning are varied, and the reach of technologies is expanding Computer vision is everywhere: Uber uses Microsoft-provided computer vision to scan driver faces and confirm their identity when they are starting a shift. The congressional television network C-SPAN uses Amazon's image recognition tools to identify on-screen lawmakers so they can quickly place a name below their image. Helping delivery drivers optimize pickup and dropoff schedules is a particularly thorny problem helped by AI. United Parcel Service estimates that for every mile that its drivers reduce their daily route, the firm saves $50 million a year

Gartner estimates that AI delivered some $1.2 trillion in aggregate business value in 2018

With AI becoming more accessible and ubiquitous, the market is exploding. IDC predicts worldwide spending on cognitive and Artificial Intelligence systems will reach $77.6B in 2022. Public companies mentioned AI and machine learning in their earnings reports more than 700 times in Q4 2017—that's seven times more than the period just two years earlier. But mentions don't mean success.

Genetic algorithms def

model-building techniques where computers examine many potential solutions to a problem, iteratively modifying (mutating) various mathematical models, and comparing the mutated models to search for a best alternative function Many computer scientists would say that neural networks approximate functions, while genetic algorithms refine functions to optimize solutions. For most managers it's useful just to know the term as a type of automated model development that's another arrow in the AI quiver. Genetic algorithms have been used for everything from building financial trading models to handling complex airport scheduling to designing parts for the international space station.

machine learning at the McDonald's Drive-thru

paid $300 million, its largest acquisition in 20 years, for the AI firm Dynamic Yield. The world's largest fast food chain serves over 68 million customers a day, just through its drive-thru lanes. AI is hoped to beef-up profits from its high-volume, low-margin business. Weather, time of day, local traffic, customer line length, wait times, nearby events, and sales data provide just some of the inputs for model building For example, if the drive-thru gets backed up, menus will switch on-the-fly to promote items that can require less prep time. API technology, which you read about in earlier chapters, made the acquisition especially attractive. McDonald's has already upgraded most US restaurants with digital signage inside and out. Since the new AI tech is accessible via API calls, integration with systems already in place will require little additional investment—just plug display software into the programming "hooks" of the API call to drive displays based on back-end smarts.

unsupervised learning def

where data are not explicitly labeled and don't have a predetermined result. Clustering customers into previously unknown groupings machine be one example

semi-supervised learning def

where data used to build models that determine an end result may contain data that has outputs explicitly labeled as well as unlabeled, e.g., "hey software, take a look at my categorizations and see if they are valid or you can come up with better or missing ones"

I Am Not a Robot, but I've Occasionally Taught Some

xxx