Fundamental Use Cases for Artificial Intelligence
In this chapter, we are going to discuss some of the use cases for Artificial Intelligence (AI). This by no means is an exhaustive list. Many industries have been impacted by AI, and the list of those industries not yet impacted gets shorter every day. However, it will be a while until we are able to replace hair stylists and plumbers. Both of these jobs require a lot of finesse and detail that robots have yet to master. I know it will be a long time before my wife trusts her hair to anyone else other than her current hair stylist, let alone a robot.
This chapter will discuss:
- Some representative AI use cases
- The jobs that will take the longest to replace by automation
- The industries that will be most impacted by AI
Representative AI use cases
From finance to medicine, it is difficult to find an industry that is not being disrupted by Artificial Intelligence. We will focus on real-world examples of the most popular applications of AI in our everyday life. We will explore the current state of the art as well as what is coming soon. Most importantly, maybe this book will spark your imagination and you will come up with some new and innovative ideas that will positively impact society and we can add it to the next edition of our book.
Artificial Intelligence, cognitive computing, machine learning, and deep learning are only some of the disruptive technologies that are enabling rapid change today. These technologies can be adopted quicker because of advances in cloud computing, Internet of Things (IoT), and edge computing. Organizations are reinventing the way they do business by cobbling together all these technologies. This is only the beginning; we are not even in the first inning, we haven't even recorded the first strike!
With that, let's begin to look at some contemporary applications of AI.
Digital personal assistants and chatbots
Unfortunately, it is still all too common for some call centers to use legacy Interactive Voice Response (IVR) systems that make calling them an exercise in patience. However, we have made great advances in the area of natural language processing: chatbots. Some of the most popular examples are:
- Google Assistant: Google Assistant was launched in 2016 and is one of the most advanced chatbots available. It can be found in a variety of appliances such as telephones, headphones, speakers, washers, TVs, and refrigerators. Nowadays, most Android phones include Google Assistant. Google Home and Nest Home Hub also support Google Assistant.
- Amazon Alexa: Alexa is a virtual assistant developed and marketed by Amazon. It can interact with users by voice and by executing commands such as playing music, creating to-do lists, setting up alarms, playing audiobooks, and answering basic questions. It can even tell you a joke or a story on demand. Alexa can also be used to control compatible smart devices. Developers can extend Alexa's capabilities by installing skills. An Alexa skill is additional functionality developed by third-party vendors.
- Apple Siri: Siri can accept user voice commands and a natural language user interface to answer questions, make suggestions, and perform actions by parsing these voice commands and delegating these requests to a set of internet services. The software can adapt to users' individual language usage, their searches, and preferences. The more it is used the more it learns and the better it gets.
- Microsoft Cortana: Cortana is another digital virtual assistant, designed and created by Microsoft. Cortana can set reminders and alarms, recognize natural voice commands, and it answers questions using information.
- Control devices in your home
- Play music and display videos on command
- Set timers and reminders
- Make appointments
- Send text and email messages
- Make phone calls
- Open applications
- Read notifications
- Perform translations
- Order from e-commerce sites
- Checking into your flight
- Booking a hotel
- Making a restaurant reservation
All these platforms also support 3rd party developers to develop their own applications or "skills" as Amazon calls them. So, the possibilities are endless.
Some examples of existing Alexa skills:
- MySomm: Recommends what wine goes with a certain meat
- The bartender: Provides instructions on how to make alcoholic drinks
- 7-minute workout: Will guide you through a tough 7-minute workout
- Uber: Allows you to order an Uber ride through Alexa
All the preceding services listed continue to get better. They continuously learn from interactions with customers. They are improved both by the developers of the services as well as by the systems taking advantage of new data points created daily by users of the services.
Most cloud providers make it extremely easy to create chatbots and for some basic examples it is not necessary to use a programming language. In addition, it is not difficult to deploy these chatbots to services such as Slack, Facebook Messenger, Skype, and WhatsApp.
Self-driving or driverless cars are vehicles that can travel along a pre-established route with no human assistance. Most self-driving cars in existence today do not rely on a single sensor and navigation method and use a variety of technologies such as radar, sonar, lidar, computer vision, and GPS.
As technologies emerge, industries start creating standards to implement and measure their progress. Driverless technologies are no different. SAE International has created standard J3016, which defines six levels of automation for cars so that automakers, suppliers, and policymakers can use the same language to classify the vehicle's level of sophistication:
Level 0 (No automation)
The car has no self-driving capabilities. The driver is fully involved and responsible. The human driver steers, brakes, accelerates, and negotiates traffic. This describes most current cars on the road today.
Level 1 (Driver assistance)
System capability: Under certain conditions, the car controls either the steering or the vehicle speed, but not both simultaneously.
Driver involvement: The driver performs all other aspects of driving and has full responsibility for monitoring the road and taking over if the assistance system fails to act appropriately. For example, Adaptive cruise control.
Level 2 (Partial automation)
The car can steer, accelerate, and brake in certain circumstances. The human driver still performs many maneuvers like interpreting and responding to traffic signals or changing lanes. The responsibility for controlling the vehicle largely falls on the driver. The manufacturer still requires the driver to be fully engaged. Examples of this level are:
- Audi Traffic Jam Assist
- Cadillac Super Cruise
- Mercedes-Benz Driver Assistance Systems
- Tesla Autopilot
- Volvo Pilot Assist
Level 3 (Conditional automation)
The pivot point between levels 2 and 3 is critical. The responsibility for controlling and monitoring the car starts to change from driver to computer at this level. Under the right conditions, the computer can control the car, including monitoring the environment. If the car encounters a scenario that it cannot handle, it requests that the driver intervene and take control. The driver normally does not control the car but must be available to take over at any time. An example of this is Audi Traffic Jam Pilot.
Level 4 (High automation)
The car does not need human involvement under most conditions but still needs human assistance under some road, weather, or geographic conditions. Under a shared car model restricted to a defined area, there may not be any human involvement. But for a privately-owned car, the driver might manage all driving duties on surface streets and the system takes over on the highway. Google's now defunct Firefly pod-car is an example of this level. It didn't have pedals or a steering wheel. It was restricted to a top speed of 25 mph and it was not used in public streets.
Level 5 (Full automation)
The driverless system can control and operate the car on any road and under any conditions that a human driver could handle. The "operator" of the car only needs to enter a destination. Nothing at this level is in production yet but a few companies are close and might be there by the time the book is published.
We'll now review some of the leading companies working in the space:
As of 2018, Waymo's autonomous cars have driven eight million miles on public roads as well as five billion miles in simulated environments. In the next few years, it is all but a certainty that we will be able to purchase a car capable of full driving autonomy. Tesla, among others, already offers driver assistance with their Autopilot feature and possibly will be the first company to offer full self-driving capabilities. Imagine a world where a child born today will never have to get a driver's license! The disruption caused in our society by this advance in AI alone will be massive. The need for delivery drivers, taxi drivers, and truckers will be obviated. Even if there are still car accidents in a driverless future, millions of lives will be saved because we will eliminate distracted driving and drunk driving.
Uber's Advanced Technology Group (ATG) is an Uber subsidiary working on developing self-driving technology. In 2016, Uber launched an experimental car service on the streets of Pittsburgh. Uber has plans to buy up to 24,000 Volvo XC90 and equip them with their self-driving technology and start commercializing them in some capacity by 2021.
Tragically, in March 2018, Elaine Herzberg was involved in an incident with an Uber driverless car and died. According to police reports, she was struck by the Uber vehicle while trying to cross the street, while she was watching a video on her phone. Ms. Herzberg became one of the first individuals to die in an incident involving a driverless car. Ideally, we would like to see no accidents ever happen with this technology, yet the level of safety that we demand needs to be tempered with the current crisis we have with traffic accidents. For context, there were 40,100 motor vehicle deaths in the US in 2017; even if we continue to see accidents with automated cars, if this death toll was slashed by say, half, thousands of lives would be saved each year.
It is certainly possible to envision a driverless vehicle that looks more like a living room than the interior of our current cars. There would be no need for steering wheels, pedals or any kind of manual control. The only input the car would need is your destination, which could be given at the beginning of your journey by "speaking" to your car. There would be no need to keep track of a maintenance schedule as the car would be able to sense when a service is due or there is an issue with the car's function.
Liability for car accidents will shift from the driver of the vehicle to the manufacturer of the vehicle doing away with the need to have car insurance. This last point is probably one of the reasons why car manufacturers have been slow to deploy this technology. Even car ownership might be flipped on its head since we could summon a car whenever we need one instead of needing one all the time.
Shipping and warehouse management
An Amazon sorting facility is one of the best examples of the symbiotic relationship that is forming between humans, computers, and robots. Computers take customer orders and decide where to route merchandise, the robots act as mules carrying the pallets and inventory around the warehouse. Humans plug the "last mile" problem by hand picking the items that are going into each order. Robots are proficient in mindlessly repeating a task many times as long as there is a pattern involved and some level of pretraining is involved to achieve this. However, having a robot pick a 20-pound package and immediately being able to grab an egg without breaking it is one of the harder robotics problems.
Robots struggle dealing with objects of different sizes, weights, shapes, and fragility; a task that many humans can perform effortlessly. People, therefore, handle the tasks that the robots encounter difficulty with. The interaction of these three types of different actors translates into a finely tuned orchestra that can deliver millions of packages everyday with very little mistakes.
Even Scott Anderson, Amazon's director of robotics fulfillment acknowledged in May 2019 that a fully automated warehouse is at least 10 years away. So, we will continue to see this configuration in warehouses across the world for a little longer.
AI can assist in generating drug candidates (that is, molecules to be tested for medical application) and then quickly eliminating some of them using constraint satisfaction or experiment simulation. We will learn more about constraint satisfaction programming in later chapters. In a nutshell, this approach allows us to speed up drug discovery by quickly generating millions of possible drug candidates and just as quickly rejecting them if the candidates do not satisfy certain predetermined constraints.
In addition, in some cases we can simulate experiments in the computer that otherwise would be much more expensive to perform in real life.
Furthermore, in some instances researchers still conduct real-world experiments but rely on robots to perform the experiments and speed up the process with them. These emerging fields are dubbed high throughput screening (HTS) and virtual high throughput screening (VHTS).
Machine learning is starting to be used more and more to enhance clinical trials. The consulting company of Accenture has developed a tool called intelligent clinical trials (ITP). It is used to predict the length of clinical trials.
Another approach that can surprisingly be used is to apply to drug discovery is Natural Language Processing (NLP). Genomic data can be represented using a string of letters and the NLP techniques can be used to process or "understand" what the genomic sequences mean.
Machine learning algorithms can be used to better price insurance by more accurately predicting how much will be spent on a patient, how good a driver an individual is, or how long a person will live.
As an example, the young.ai project from Insilico Medicine can predict with some accuracy how long someone will live from a blood sample and a photograph. The blood sample provides 21 biomarkers such as cholesterol level, inflammation markers, hemoglobin counts and albumin level that are used as input to a machine learning model. Other inputs into the model are ethnicity and age, as well as a photograph of the person.
Interestingly, as of now, anyone can use this service for free by visiting young.ai (https://young.ai) and providing the required information.
Doctors can make better diagnosis on their patients and be more productive in their practice by using sophisticated rules engines and machine learning. As an example, in a recent study at the University of California in San Diego conducted by Kang Zhang , one system could diagnose children's illnesses with a higher degree of accuracy than junior pediatricians. The system was able to diagnose the following diseases with a degree of accuracy of between 90% and 97%:
- Glandular fever
- Chicken pox
- Hand, foot, and mouth disease
Medical imaging interpretation
Medical imaging data is a complex and rich source of information about patients. CAT scans, MRIs, and X-rays contain information that is otherwise unavailable. There is a shortage of radiologists and clinicians that can interpret them. Getting results from these images can sometimes take days and can sometimes be misinterpreted. Recent studies have found that machine learning models can perform just as well, if not better, than their human counterparts.
Data scientists have developed AI enabled platforms that can interpret MRI scans and radiological images in a matter of minutes instead of days and with a higher degree of accuracy when compared with traditional methods.
Perhaps surprisingly, far from being concerned, leaders from the American College for Radiology see the advent of AI as a valuable tool for physicians. In order to foster further development in the field, the American College for Radiology Data Science Institute (ACR DSI) released several AI use cases in medical imaging and plans to continue releasing more.
An hour-long session with a psychiatrist can costs hundreds of dollars. We are on the cusp of being able to simulate the behavior with AI chatbots. At the very least, these bots will be able to offer follow-up care from the sessions with the psychiatrist and help with a patient's care between doctor's visits.
One early example of an automated counselor is Eliza. It was developed in 1966 by Joseph Weizenbaum. It allows users to have a "conversation" with the computer mimicking a Rogerian psychotherapist. Remarkably, Eliza feels natural, but its code is only a few hundred lines and it doesn't really use much AI at its core.
A more recent and advanced example is Ellie. Ellie was created by the Institute for Creative Technologies at the University of Southern California. It helps with the treatment of people with depression or post-traumatic stress disorder. Ellie is a virtual therapist (she appears on screen), responds to emotional cues, nods affirmatively when appropriate and shifts in her seat. She can sense 66 points on a person's face and use these inputs to read a person's emotional state. One of Ellie's secrets is that she is obviously not human and that makes people feel less judged and more comfortable opening up to her.
Smart health records
Medicine is notorious for being a laggard in moving to electronic records. Data science provides a variety of methods to streamline the capture of patient data including OCR, handwriting recognition, voice to text capture, and real-time reading and analysis of patient's vital signs. It is not hard to imagine a future coming soon where this information can be analyzed in real-time by AI engines to take decisions such as adjusting body glucose levels, administering a medicine, or summoning medical help because a health problem is imminent.
Disease detection and prediction
The human genome is the ultimate dataset. At some point soon, we will be able to use the human genome as input to machine learning models and be able to detect and predict a wide variety of diseases and conditions using this vast dataset.
Using genomic datasets as an input in machine learning is an exciting area that is evolving rapidly and will revolutionize medicine and health care.
The human genome contains over 3 billion base pairs. We are making progress on two fronts that will accelerate progress:
- Continuous advancements in the understanding of genome biology
- Advances in big data computing to process vast amounts of data faster
There is much research applying deep learning to the field of genomics. Although it is still in early stages, deep learning in genomics has the potential to inform fields including:
- Functional genomics
- Population genetics
- Clinical genetics
- Crop yield improvement
- Epidemiology and public health
- Evolutionary and phylogenetic analysis
We have gotten to a point where, in some cases, we don't even realize we are using artificial intelligence. A sign that a technology or product is good is when we don't necessarily stop to think how it's doing what it is doing. A perfect example of this is Google Search. The product has become ubiquitous in our lives and we don't realize how much it relies on artificial intelligence to produce its amazing results. From its Google Suggest technology to its constant improvement of the relevancy of its results, AI is deeply embedded in its search process.
Early in 2015, as was reported by Bloomberg, Google began using a deep learning system called RankBrain to assist in generating search query responses. The Bloomberg article describes RankBrain as follows:
"RankBrain uses artificial intelligence to embed vast amounts of written language into mathematical entities — called vectors — that the computer can understand. If RankBrain sees a word or phrase it isn't familiar with, the machine can make a guess as to what words or phrases might have a similar meaning and filter the result accordingly, making it more effective at handling never-before-seen search queries."
— Clark, Jack 
As of the last report, RankBrain plays a role in a large percentage of the billions of Google Search queries. As one can imagine, the company is tight lipped about how exactly RankBrain works, and furthermore even Google might have a hard time explaining how it works. You see, this is one of the dilemmas of deep learning. In many cases, it can provide highly accurate results, but deep learning algorithms are usually hard to understand in terms of why an individual answer was given. Rule-based systems and even other machine learning models (such as Random Forest) are much easier to interpret.
The lack of explainability of deep learning algorithms has major implications, including legal implications. Lately, Google and Facebook among others, have found themselves under the microscope to determine if their results are biased. In the future, legislators and regulators might require that these tech giants provide a justification for a certain result. If deep learning algorithms do not provide explainability, they might be forced to use other less accurate algorithms that do.
Initially, RankBrain only assisted in about 15 percent of Google queries, but now it is involved in almost all user queries.
However, if a query is a common query, or something that the algorithm understands, the RankBrain rank score is given little weight. If the query is one that the algorithm has not seen before or it does not know its meaning, RankBrain score is much more relevant.
Recommendation systems are another example of AI technology that has been weaved into our everyday lives. Amazon, YouTube, Netflix, LinkedIn, and Facebook all rely on recommendation technology and we don't even realize we are using it. Recommendation systems rely heavily on data and the more data that is at their disposable, the more powerful they become. It is not coincidence that these companies have some of the biggest market caps in the world and their power comes from them being able to harness the hidden power in their customer's data. Expect this trend to continue in the future.
What is a recommendation? Let's answer the question by first exploring what it is not. It is not a definitive answer. Certain questions like "what is two plus two?" or "how many moons does Saturn have?" have a definite answer and there is no room for subjectivity. Other questions like "what is your favorite movie?" or "do you like radishes?" are completely subjective and the answer is going to depend on the person answering the question. Some machine learning algorithms thrive with this kind of "fuzziness." Again, these recommendations can have tremendous implications.
Think of the consequences of Amazon constantly recommending a product versus another. The company that makes the recommended product will thrive and the company that makes the product that was not recommended could go out of business if it doesn't find alternative ways to distribute and sell its product.
One of the ways that a recommender system can improve is by having previous selections from users of the system. If you visit an e-commerce site for the first time and you don't have an order history, the site will have a hard time making a recommendation tailored to you. If you purchase sneakers, the website now has one data point that it can start using as a starting point. Depending on the sophistication of the system, it might recommend a different pair of sneakers, a pair of athletic socks, or maybe even a basketball (if the shoes were high-tops).
An important component of good recommendation systems is a randomization factor that occasionally "goes out on a limb" and makes oddball recommendations that might not be that related to the initial user's choices. Recommender systems don't just learn from history to find similar recommendations, but they also attempt to make new recommendations that might not be related at first blush. For example, a Netflix user might watch "The Godfather" and Netflix might start recommending Al Pacino movies or mobster movies. But it might recommend "Bourne Identity," which is a stretch. If the user does not take the recommendation or does not watch the movie, the algorithm will learn from this and avoid other movies like the "Bourne Identity" (for example any movies that have Jason Bourne as the main character).
As recommender systems get better, the possibilities are exciting. They will be able to power personal digital assistants and become your personal butler that has intimate knowledge of your likes and dislikes and can make great suggestions that you might have not thought about. Some of the areas where recommendations can benefit from these systems are:
- Potential partners (online dating)
- Books and articles
- Search results
- Financial services (robo-advisors)
Some notable specific examples of recommender systems follow:
A contest that created a lot of buzz in the recommender system community was the Netflix Prize. From 2006 to 2009, Netflix sponsored a competition with a grand prize of one million US dollars. Netflix made available a dataset of 100 million plus ratings.
Netflix offered to pay the prize to the team that offered the highest accuracy in their recommendations and was 10% more accurate than the recommendations from Netflix's existing recommender system. The competition energized research for new and more accurate algorithms. In September 2009, the grand prize was awarded to the BellKor's Pragmatic Chaos team.
Pandora is one of the leading music services. Unlike other companies like Apple and Amazon, Pandora's exclusive focus is as a music service. One of Pandora's salient service features is the concept of customized radio stations. These "stations" allow users to play music by genre. As you can imagine, recommender systems are at the core of this functionality.
Pandora's recommender is built on multiple tiers:
- First, their team of music experts annotates songs based on genre, rhythm, and progression.
- These annotations are transformed into a vector for comparing song similarity. This approach promotes the presentation of "long tail" or obscure music from unknown artists that nonetheless could be a good fit for individual listeners.
- The service also heavily relies on user feedback and uses it to continuously enhance the service. Pandora has collected over 75 billion feedback data points on listener preferences.
- The Pandora recommendation engine can then perform personalized filtering based on a listener's preferences using their previous selections, geography, and other demographic data.
In total, Pandora's recommender uses around 70 different algorithms, including 10 to analyze content, 40 to process collective intelligence, and about another 30 to do personalized filtering.
Robo-advisors are recommendation engines that provide investment or financial advice and management with minimal human involvement. These services use machine learning to automatically allocate, manage, and optimize a customer's asset mix. They can offer these services at a lower cost than traditional advisors because their overhead is lower, and their approach is more scalable.
There is now fierce competition in this space with well over 100 companies offering these kinds of services. Robo-advisors are considered a tremendous breakthrough. Formerly, wealth management services were an exclusive and expensive service reserved for high net worth individuals. Robo-advisors promise to bring a similar service to a broader audience with lower costs compared to the traditional human-enabled services. Robo-advisors could potentially allocate investments in a wide variety of investment products like stocks, bonds, futures, commodities, real estate, and other exotic investments. However, to keep things simple investments are often constrained to exchange traded funds (ETFs).
As we mentioned there are many companies offering robo-advice. As an example, you might want to investigate Betterment to learn more about this topic. After filling out a risk questionnaire, Betterment will provide users with a customized, diversified portfolio. Betterment will normally recommend a mix of low-fee stock and bond index funds. Betterment charges an administration fee (as a percentage of the portfolio) but it is lower than most human-powered services. Please note that we are not endorsing this service and we only mention it as an example of a recommendation engine in the financial sector.
The smart home
Whenever you bring up the topic of AI to the common folk on the street, they are usually skeptical about how soon it is going to replace human workers. They can rightly point to the fact that we still need to do a lot of housework around the house. AI needs to become not only technologically possible, but it also needs to be economically feasible for adoption to become widespread. House help is normally a low-wage profession and, for that reason, automation to replace it needs to be the same price or cheaper. In addition, house work requires a lot of finesse and it comprises tasks that are not necessarily repetitive. Let's list out some of the tasks that this automaton will need to perform in order to be proficient:
- Wash and dry clothes
- Fold clothes
- Cook dinner
- Make beds
- Pick up items off the floor
- Mop, dust and vacuum
- Wash dishes
- Monitor the home
As we already know, some of these tasks are easy to perform for machines (even without AI) and some of them are extremely hard. For this reason and because of the economic considerations, the home will probably be one of the last places to become fully automated. Nonetheless, let's look at some of the amazing advances that have been made in this area.
Home monitoring is one area where great solutions are generally available already. The Ring video doorbell from Amazon and the Google Nest thermostat are two inexpensive options that are widely available and popular. These are two simple examples of smart home devices that are available for purchase today.
The Ring video doorbell is a smart home device connected to the internet that can notify the homeowner of activity at their home, such as a visitor, via their smartphone. The system does not continuously record but rather it activates when the doorbell is pressed, or when the motion detector is activated. The Ring doorbell can then let the home owner watch the activity or communicate with the visitor using the built-in microphone and speakers. Some models also allow the homeowner to open the door remotely via a smart lock and let the visitor into the house.
The Nest Learning Thermostat is a smart home device initially developed by Nest Labs, a company that was later bought by Google. It was designed by Tony Fadell, Ben Filson, and Fred Bould. It is programmable, Wi-Fi-enabled, and self-learning. It uses artificial intelligence to optimize the temperature of the home while saving energy.
In the first weeks of use you set the thermostat to your preferred settings and this will serve as a baseline. The thermostat will learn your schedule and your preferred temperatures. Using built-in sensors and your phones' locations, the thermostat will shift into energy saving mode when no one is home.
Since 2011, the Nest Thermostat has saved billions of kWh of energy in millions of homes worldwide. Independent studies have shown that it saves people an average of 10% to 12% on their heating bills and 15% on their cooling bills so in about 2 years it may pay for itself.
Vacuuming and mopping
Two tasks that have been popular to hand off to robots are vacuuming and mopping. A robotic vacuum cleaner is an autonomous robotic vacuum cleaner that uses AI to vacuum a surface. Depending on the design, some of these machines use spinning brushes to reach tight corners and some models include several other features in addition to being able to vacuum, such as mopping and UV sterilization. Much of the credit for popularizing this technology goes to the company (not the film), iRobot.
iRobot was started in 1990 by Rodney Brooks, Colin Angle, and Helen Greiner after meeting each other while working in MIT's Artificial Intelligence Lab. iRobot is best known for its vacuuming robot (Roomba), but for a long time they also had a division devoted to the development of military robots. The Roomba started selling in 2002. As of 2012 iRobot had sold more than eight million home robots as well as creating more than 5,000 defense and security robots. The company's PackBot is a bomb-disposal robot used by the US military that has been used extensively in Iraq and Afghanistan. PackBots were also used to gather information under dangerous conditions at the Fukushima Daiichi nuclear disaster site. iRobot's Seaglider was used to detect underwater pools of oil after the Deepwater Horizon oil spill in the Gulf of Mexico.
Another iRobot product is the Braava series of cleaners. The Braava is a small robot that can mop and sweep floors. It is meant for small spaces like bathrooms and kitchens. It sprays water and uses an assortment of different pads to clean effectively and quietly. Some of the Braava models have a built-in navigation system. The Braava doesn't have enough power to remove deep-set stains, so it's not a complete human replacement, but it does have wide acceptance and high ratings. We expect them to continue to gain popularity.
The potential market for intelligent devices in the home is huge and it is all but certain that we will continue to see attempts from well established companies and startups alike to exploit this largely untapped market.
Picking up your mess
As we learned in the shipping use case, picking objects of different weights, dimensions, and shapes is one of the most difficult tasks to automate. Robots can perform efficiently under homogeneous conditions like a factory floor where certain robots specialize in certain tasks. Picking up a pair of shoes after picking up a chair, however, can be immensely challenging and expensive. For this reason, do not expect this home chore to be pervasively performed by machines in a cost-effective fashion any time soon.
- Certain restaurants may charge hundreds of dollars for their food and be paying high prices for skilled chefs. Therefore, they might be open to using technology to replace their high-priced staff if this should work out to be more profitable. An example for this is a five-star sushi restaurant.
- Some tasks in the kitchen are repetitive and therefore lend themselves to automation. Think of a fast food joint where hamburgers and fries might have to be made by the hundreds. Thus, rather than having one machine handle the entire disparate cooking process, a series of machines could deal with individual repetitive stages of the process.
Smart prosthetics are great examples of artificial intelligence augmenting humans rather than replacing them. There are more than a few chefs that lost their arm in an accident or were born without a limb.
One example is chef Michael Caines who runs a two Michelin star restaurant and lost his arm in a horrific car accident. Chef Caines was head chef of Gidleigh Park in Devon in England until January 2016. He is currently the executive chef of the Lympstone Manor hotel between Exeter and Exmouth. He now cooks with a prosthetic arm, but you'd never know it given the quality of his food.
Another example is Eduardo Garcia who is a sportsman and a chef – both of which are made possible by the most advanced bionic hand in the world.
On October 2011, while bow-hunting elk he was electrocuted in the Montana backcountry. Eduardo was hunting by himself in October 2011. He was in back country when he saw a dead baby black bear. He stopped to check it out, knelt, and used his knife to prod it.
While doing so, 2,400 volts coursed through his body – the baby bear had been killed by a buried, live electrical wire. He survived but lost his arm during the incident.
In September 2013, Garcia was fitted by Advanced Arm Dynamics with a bionic hand designed by Touch Bionics. The bionic hand is controlled by Garcia's forearm muscles and can grip in 25 different ways. With his new hand, Garcia can perform tasks that normally require great dexterity. His new hand still has some limitations. For example, Garcia cannot lift heavy weights. However, there are things that he can perform now that he couldn't before. For example, he can grab things out of a hot oven and not get burnt and it is impossible to cut his fingers.
Conversely, rather than augmenting humans, robots may replace humans in the kitchen entirely. An example of this is Moley, the robotic kitchen. Moley is not currently in production but the most advanced prototype of the Moley Robotic Kitchen consists of two robotic arms with hands equipped with tactile sensors, a stove top, an oven, a dishwasher, and a touchscreen unit. These artificial hands can lift, grab, and interact with most kitchen equipment including knives, whisks, spoons, and blenders.
Using a 3D camera and a glove it can record a human chef preparing a meal and then upload detailed steps and instructions into a repository. The chef's actions are then translated into robotic movements using gesture recognition models. These models were created in collaboration with Stanford University and Carnegie Mellon University. After that Moley can reproduce the same steps and cooks the exact same meal from scratch.
In the current prototype, the user can operate it using a touchscreen or smartphone application with ingredients prepared in advance and placed in preset locations. The company's long-term goal is to allow users to simply select an option from a list of more 2,000 recipes and Moley will have the meal prepared in minutes.
There is perhaps no better example to demonstrate the awe-inspiring advances in Artificial Intelligence than the progress that has been made in the area of gaming. Humans are competitive by nature and having machines beat us at our own games is an interesting yardstick to measure the breakthroughs in the field. Computers have long been able to beat us in some of the more basic, more deterministic, less compute-intensive games like say checkers. It's only in the last few years that machines have been able to consistently beat the masters of some of the harder games. In this section we go over three of these examples.
Video games have been used for decades as a benchmark to test the performance of AI systems. As capabilities increase, researchers work with more complex games that require different types of intelligence. The strategies and techniques developed from this game playing can transfer to solving real-world problems. The game of StarCraft II is considered one of the hardest, though it is an ancient game by video game standards.
The team at DeepMind introduced a program dubbed AlphaStar that can play StarCraft II and was for the first time able to defeat a top professional player. In matches held in December 2018, AlphaStar whooped a team put together by Grzegorz "MaNa" Komincz, one of the world's strongest professional StarCraft players with a score of 5-0. The games took place under professional match conditions and without any game restrictions.
In contrast to previous attempts to master the game using AI that required restrictions, AlphaStar can play the full game with no restrictions. It uses a deep neural network that is trained directly from raw game data using supervised learning and reinforcement learning.
One of the things that makes StarCraft II so difficult is the need to balance short-and long-term goals and adapt to unexpected scenarios. This has normally posed a tremendous challenge for previous systems.
While StarCraft is just a game, albeit a difficult one, the concepts and techniques coming out of AlphaStar can be useful in solving other real-world challenges. As an example, AlphaStar's architecture is capable of modeling very long sequences of likely actions – with games often lasting up to an hour with tens of thousands of moves – based on imperfect information. The primary concept of making complicated predictions over long sequences of data can be found in many real-world problems, such as:
- Weather prediction
- Climate modelling
- Natural Language Understanding
The success that AlphaStar has demonstrated playing StarCraft represents a major scientific breakthrough in one of the hardest video games in existence. These breakthroughs represent a big leap in the creation of artificial intelligence systems that can be transferred and that can help solve fundamental real-world practical problems.
Ken Jennings has the longest unbeaten run in the show's history with 74 consecutive appearances. Brad Rutter had the distinction of winning the biggest prize pot with a total of $3.25 million.
Both players agreed to an exhibition match against Watson.
Watson is a question-answering system that can answer questions posed in natural language. It was initially created by IBM's DeepQA research team, led by principal investigator David Ferrucci.
The main difference between the question-answering technology used by Watson and general search (think Google searches) is that general search takes a keyword as input and responds with a list of documents with a ranking based on the relevance to the query. Question-answering technology like what is used by Watson takes a question expressed in natural language, tries to understand the question at a deeper level, and tries to provide the precise answer to the question.
The software architecture of Watson uses:
- IBM's DeepQA software
- Apache UIMA (Unstructured Information Management Architecture)
- A variety of languages, including Java, C++, and Prolog
- SUSE Linux Enterprise Server
- Apache Hadoop for distributed computing
In the first round of play Deep Blue won the first game against Gary Kasparov. However, they were scheduled to play six games. Kasparov won three and drew two of the following five games thus defeating Deep Blue by a score of 4–2.
The Deep Blue team went back to the drawing board, made a lot of enhancements to the software, and played Kasparov again in 1997. Deep Blue won the second round against Kasparov winning the six-game rematch by a score of 3½–2½. It then became the first computer system to beat a current world champion in a match under standard chess tournament rules and time controls.
A lesser known example, and a sign that machines beating humans is becoming common place, is the achievement in the area of chess by the AlphaZero team.
Google scientists from their AlphaZero research team created a system in 2017 that took just four hours to learn the rules of chess before crushing the most advanced world champion chess program at the time called Stockfish. By now the question as to whether computers or humans are better at chess has been resolved.
Let's pause for a second and think about this. All of humanity's knowledge about the ancient game of chess was surpassed by a system that, if it started learning in the morning, would be done by lunch time.
The system was given the rules of chess, but it was not given any strategies or further knowledge. Then, in a few hours, AlphaZero mastered the game to the extent it was able to beat Stockfish.
In a series of 100 games against Stockfish, AlphaZero won 25 games while playing as white (white has an advantage because it goes first). It also won three games playing as black. The rest of the games were ties. Stockfish did not obtain a single win.
Not only are there more possible (19 x 19) Go-board positions than there are atoms in the visible universe and the number of possible chess positions is negligible to the number of Go positions. But Go is at least several orders of magnitude more complex than a game of chess because of the large number of possible ways to let the game flow with each move towards another line of development. With Go, the number of moves in which a single stone can affect and impact the whole-board situation is also many orders of magnitude larger than that of a single piece movement with chess.
There is great example of a powerful program that can play the game of Go also developed by DeepMind called AlphaGo. AlphaGo also has three far more powerful successors, called AlphaGo Master, AlphaGo Zero, and AlphaZero.
In October 2015, the original AlphaGo became the first computer Go program to beat a human professional Go player without handicaps on a full-sized 19 x 19 board. In March 2016, it beat Lee Sedol in a five-game match. This became the first time a Go program beat a 9-dan professional without handicaps. Although AlphaGo lost to Lee Sedol in the fourth game, Lee resigned in the final game, giving a final score of 4 games to 1.
At the 2017 Future of Go Summit, the successor to AlphaGo called AlphaGo Master beat the master Ke Jie in a three-game match. Ke Jie was ranked the world No.1 ranked player at the time. After this, AlphaGo was awarded professional 9-dan by the Chinese Weiqi Association.
AlphaGo and its successors use a Monte Carlo tree search algorithm to find their moves based on knowledge previously "learned" by machine learning, specifically using deep learning and training, both playing with humans and by itself. The model is trained to predict AlphaGo's own moves and the winner's games. This neural net improves the strength of tree search, resulting in better moves and stronger play in following games.
It is all but a certainty that within the next few decades it will be possible to create movies that are 100% computer generated. It is not unfathomable to envision a system where the input is a written script and the output is a full-length feature film. In addition, some strides have been made in natural generators. So, eventually not even the script will be needed. Let's explore this further.
A deepfake is a portmanteau, or blend, of "deep learning" and "fake." It is an AI technique to merge video images. A common application is to overlap someone's face onto another. A nefarious version of this was used to merge pornographic scenes with famous people or to create revenge porn. Deepfakes can also be used to create fake news or hoaxes. As you can imagine, there are severe societal implications if this technology is misused.
One recent version of similar software was developed by a Chinese company called Momo who developed an app called Zao. It allows you to overlap someone's face over short movie clips like Titanic and the results are impressive. This and other similar applications do not come without controversy. Privacy groups are complaining that the photos submitted to the site per the terms of the user agreement become property of Momo and then can later be used for other applications.
It will be interesting to see how technology continues to advance in this area.
Movie Script Generation
Sunspring is an experimental science fiction short film released in 2016. It was entirely written by using deep learning techniques. The film's script was created using a long short-term memory (LSTM) model dubbed Benjamin. Its creators are BAFTA-nominated filmmaker Oscar Sharp and NYU AI researcher Ross Goodwin. The actors in the film are Thomas Middleditch, Elisabeth Grey, and Humphrey Ker. Their character names are H, H2, and C, living in the future. They eventually connect with each other and a love triangle forms.
Underwriting and deal analysis
What is underwriting? In short, underwriting is the process by which an institution determines if they want to take a financial risk in exchange for a premium. Examples of transactions that require underwriting are:
- Issuing an insurance policy
- Installment loans
- Credit cards
- Commercial lines of credit
- Securities underwriting and Initial Public Offerings (IPOs)
As can be expected, determining whether an insurance policy or a loan should be issued and at what price can be very costly if the wrong decision is made. For example, if a bank issues a loan and the loan defaults, it would require dozens of other performing loans to make up for that loss. Inversely, if the bank passes up on a loan where the borrower was going to make all their payments is also detrimental to the bank finances. For this reason, the bank spends considerable time analyzing or "underwriting" the loan to determine the credit worthiness of the borrower as well as the value of the collateral securing the loan.
Even with all these checks, underwriters still get it wrong and issue loans that default or bypass deserving borrowers. The current underwriting process follows a set of criteria that must be met but specially for smaller banks there is still a degree of human subjectivity in the process. This is not necessarily a bad thing. Let's visit a scenario to explore this further:
A high net worth individual recently came back from a tour around the world. Three months ago, they got a job at a prestigious medical institution and their credit score is above 800.
Would you lend money to this individual? With the characteristics given, they seem to be a good credit risk. However, normal underwriting rules might disqualify them because they haven't been employed for the last two years. Manual underwriting would look at the whole picture and probably approve them.
Similarly, a machine learning model would probably be able to flag this as a worthy account and issue the loan. Machine learning models don't have hard and fast rules but rather "learn by example."
Many lenders are already using machine learning in their underwriting. An interesting example of a company that specializes in this space is Zest Finance. Zest Finance uses AI techniques to assist lenders with their underwriting. AI can help to increase revenue and reduce risk. Most importantly well applied AI in general and Zest Finance in particular can help companies to ensure that the AI models used are compliant with a country's regulations. Some AI models can be a "black box" where it is difficult to explain why one borrower was rejected and another one was accepted. Zest Finance can fully explain data modeling results, measure business impact, and comply with regulatory requirements. One of Zest Finance's secret weapons is the use of non-traditional data, including data that a lender might have in-house, such as:
- Customer support data
- Payment histories
- Purchase transactions
They might also consider nontraditional credit variables such as:
- The way a customer fills out a form
- The method a customer uses to arrive at the site or how they navigate the site
- The amount of time taken to fill out an application
Data cleansing and transformation
Just as gas powers a car, data is the lifeblood of AI. The age-old adage of "garbage in, garbage out" remains painfully true. For this reason, having clean and accurate data is paramount to producing consistent, reproducible, and accurate AI models. Some of this data cleansing has required painstaking human involvement. By some measures, it is said that a data scientist spends about 80% of their time cleaning, preparing, and transforming their input data and 20% of the time running and optimizing their models. Examples of this are the ImageNet and MS-COCO image datasets. Both contain over a million labeled images of various objects and categories. These datasets are used to train models that can distinguish between different categories and object types. Initially, these datasets were painstakingly and patiently labeled by humans. As these systems become more prevalent, we can use AI to perform the labeling. Furthermore, there is a plethora of AI-enabled tools that help with the cleansing and deduplication process.
One good example is Amazon Lake Formation. In August 2019, Amazon made its service Lake Formation generally available. Amazon Lake Formation automates some of the steps typically involved in the creation of a data lake including the collection, cleansing, deduplication, cataloging, and publication of data. The data then can be made available for analytics and to build machine models. To use Lake Formation, a user can bring data into the lake from a range of sources using predefined templates. They can then define policies that govern data access depending on the level of access that groups across the organization require.
Lake Formation also provides a centralized dashboard where administrators can manage and monitor data access policies, governance, and auditing across multiple analytics engines. Users can also search for datasets in the resulting catalog. As the tool evolves in the next few months and years, it will facilitate the analysis of data using their favorite analytics and machine learning services, including:
- Amazon Redshift
- Amazon Athena
- AWS Glue
- Amazon EMR
- Amazon QuickSight
- Amazon SageMaker
This chapter provided a few examples of the applications of AI. That said, the content here doesn't begin to scratch the surface! We tried to keep the use cases to either technology that is widely available, or at least that has the potential to become available soon. It is not difficult to extrapolate how this technology is going to continue to improve, become cheaper, and be more widely available. For example, it will be quite exciting when self-driving cars start becoming popular.
However, we can all be certain that the bigger applications of AI have not yet even been conceived. Also, advances in AI will have wide implications for our society and at some point, we will have to deal with these questions:
- What happens if an AI became so evolved that it became conscious? Should it be given rights?
- If a robot replaces a human, should companies be required to continue paying payroll tax for that displaced worker?
- Will we get to a point where computers are doing everything, and if so, how will we adapt to this; how will we spend our time?
- Worse yet, does the technology enable a few individuals to control all resources? Will a universal income society emerge in which individuals can pursue their own interests? Or will the displaced masses live in poverty?
Bill Gates and Elon Musk have warned about AIs either destroying the planet in a frenzied pursuit of their own goals or doing away with humans by accident (or not so much by accident). We will take a more optimistic "half-full" view of the impact of AI, but one thing that is certain is that it will be an interesting journey.
- Willingham, Emily, A Machine Gets High Marks for Diagnosing Sick Children, Scientific American, October 7th, 2019, https://www.scientificamerican.com/article/a-machine-gets-high-marks-for-diagnosing-sick-children/
- Clark, Jack, Google Turning Its Lucrative Web Search Over to AI Machines, Bloomberg, October 26th, 2015, https://www.bloomberg.com/news/articles/2015-10-26/google-turning-its-lucrative-web-search-over-to-ai-machines