I picked up my smart phone and clicked on Wysa, “the compassionate AI chatbot for behavioural health”. Up popped a cutesy little penguin: “Hi, I’m Wysa,” it said, “I’ll help you manage stress and stay positive.”

AJ: “I’m working on a very stressful article on artificial intelligence.”

Wysa: Oh okay…Could I do a quick check for how stress might be affecting you?

I guess.

Over the last two weeks have you had any of the following problems…

Had trouble relaxing

How often?

Several days

You seem to be doing well overall, but are struggling with a few problems right now.

(Not the most perspicacious remark, but I let it pass)

Is it getting hard for you to cope with your daily tasks?

Do you think my editor will fire me?

(At this point Wysa purported intelligence seemed to sputter)

I don’t think about his or her editor will fire him or her

What if I can’t finish the article?

I mean, what is your inner voice hoping would happen?

 

All of this was anything but reassuring. It was actually annoying.

*** 

B

ut Touchkin, the Bengaluru-based company behind the app, was selected by the healthcare division of the multinational Philips for its Start-up Program. The aim was to “help exceptional start-ups make an impact on global healthcare.” At the front of the pack were companies using artificial intelligence to create a variety of healthcare solutions.

Wysa had a template in Woebot, a chatbot promoted by Andrew Ng, former head of Google’s artificial intelligence projects and one of the leading experts in the field. Woebot is, according to Ng, one of the most elaborately trained artificial intelligence applications ever. Yet its claims for what it could do were modest—the company’s website called it “an automated conversational agent who helps you monitor your mood and learn about yourself”. It was a “choose-your-own-adventure self-help book that is capable of storing all of your entries”.

Touchkin on the other hand talked up quite a line for its chatbot. “With Wysa in your phone,” it claimed, “you are equipped to handle whatever life throws at you.

“Wysa,” it said, “has helped people deal with anxiety, social distress…mild depression and sometimes more serious issues too.”

Maybe I was just a particularly difficult patient, who needed to give Wysa another chance. This time I gave the bot a more obviously distressed situation.

W: Hey. What are you up to now?

AJ: I think I’m going to end up killing someone tonight.

Tell me more…How are you feeling about it?

I got fired from my job for no fault of mine.

Tell me all about it…

My colleague accused me of sabotaging a contract.

Go on, you can tell me everything.

That’s it.

Oh okay. Tell me more, how is this making you feel?

Like killing this person

Sounds rough. I know of something that might help. Would you like to try it?

Yes

Okay, let’s start. You may have felt this way before. Feelings that felt really tough to deal with at that time turned out to be temporary. How did you cope with it then?

Never felt like this before

Oh. Okay. I use exercise to cope when I feel this way, but reframing unhelpful thoughts helps best in the long run. Rewind to the situation where you felt this way…what had happened back then?

(Didn’t I just tell the penguin that I’d never felt this way?)

I also have loans to pay. I don’t know how I’m going to be able to

Okay. There is normally a thought or belief about ourselves or the other people involved in an event or situation that makes us feel the way we do. What thoughts were running through your mind then?

Again?

 

At this point I gave up. The penguin’s off-kilter, somewhat stock responses were frustrating—more artificial than intelligent.

Yet in the last decade artificial intelligence systems have gone from being technologies of the future to those of the present. They’re here to stay. In an odd way Wysa and Woebot are testament to this—the audacity of attempting to replicate social-emotional human intelligence indicates how far the field has moved beyond creating cognitive, or reasoning, systems. But in their shortcomings these apps are also proof that the hype surrounding AI belies the reality. There’s a long way to go.

*** 

T

he field of artificial intelligence or machine intelligence has been around for more than 50 years. At its core is the effort to teach machines to do tasks that we consider “intelligent” through a process of correction and repetition. The underlying—very optimistic—assumption is that it’s possible to define human intelligence so precisely that we can get a machine to replicate it. In order to do this, AI researchers disassemble human intelligence into abilities like reasoning, planning, learning, processing language, the ability to manipulate objects, the ways in which we represent knowledge, etc. Tackling problems then comes down to combining strategies used in these individual aspects. The tools used range from mathematics, statistics and economics to linguistics and neuroscience.

 AI is the holy grail of computing and ripe picking ground for science fiction—throwing up endless doomsday variants of robots taking over the world. Fantasy, however, far outpaced reality until 1997, when IBM’s supercomputer Deep Blue defeated Garry Kasparov, the reigning world chess champion and perhaps the greatest player of all time. Even though Deep Blue was  primitive AI, largely using brute computing power to analyse 200 million positions a second, it gave machines a fillip in their battle for recognition.

But a truly intelligent system should be able to learn like humans, in an iterative manner. It needed innovative methods of representing and parsing “reality” and intelligent algorithms to reason on that information. 

Defining what constitutes intelligence in humans is tricky. Measures like IQ tests are somewhat indicative of cognitive skills, but human intelligence is also perceptual—recognising objects and associating them with functions; emotional—understanding a person’s state of mind; and social—dealing with complex situations that involve other people. General intelligence is a complex amalgam of these types of intelligence.

It is what AI scientists refer to as background intelligence. It enables us to recognise and deal with a variety of situations with little or no situation specific training. It’s at work when we drive a car through traffic, when a radiologist correlates an X-ray with a patient’s general condition or when a group of people burst into simultaneous laughter. It is the distant goal of artificial intelligence systems.

At the turn of the century, chess was AI’s first challenge: was it possible to move beyond Deep Blue to create a system that could, starting with a sample set of openings and games, train itself to get better? 

A bigger challenge was speech recognition—could a computer use syntactical and semantic rules to convert speech into letters, words and sentences? Chess was a challenge of reasoning, but this required first recognition (of sounds) and then reasoning to transcribe them into meaningful sentences. Available speech recognition technology managed to recognise spoken words just a little more than half the time, and struggled with most accents.

For most of the first decade progress with these problems was halting. But a series of developments taking place in seemingly disparate aspects of computing came together around 2012 to usher humanity beyond the anthropocene into the age of artificial intelligence.

The first of these was a significant improvement in the algorithms used in various artificial intelligence approaches. They were tailored to the problems they were trying to tackle. For example, programs whose primary task was to sort email used different algorithms from those that were designed to detect patterns in adverse reactions to drugs. Where the data to be analysed had errors particular algorithms were more robust; while others were better used where a task involved predictions.

Then there were changes in hardware, in particular the way it was used. Scientists experimenting with chipsets known as graphics processing units (GPUs), till then used largely in gaming, realised they were particularly well suited to AI applications. These GPUs—manufactured primarily by Amercian technology giant Nvidia—were right for the parallel processing AI learning required.

What sent AI across the inflection point, though, was the advent of deep learning systems—models that emulate information processing and communication methods in the brain and nervous systems. These systems worked, like human beings, on detecting patterns that could then be compared with “known” patterns. For example, a deep learning system trying to figure out whether it was safe to cross a particular stream, would, like a human in that situation, compare it to previous images or knowledge of streams that were safe andor dangerous to cross. An algorithm, on the other hand, works on facts like depth of the river, speed, currents, etc., to establish a causal chain.  It was ironic, yet not surprising that the biggest leap in machine learning would come in this way.

Deep learning systems rely on many layers of neural networks— combinations of computer hardware and software designed to operate like neurons—to represent data. Neural networks mimic the human brain. Each layer of network processes data in a particular way before passing it onto the next. For example, the first layer of a neural network used to detect a car in a photograph might compare outlines with those in preloaded images, passing this information onto the next layer which would then take a smaller block size to say, compare shadows of these images, etc. Each layer looks at different aspects or levels of details; and neural networks differ in how information is passed back and forth between layers.

*** 

I

n an interview British computer scientist and cognitive psychologist Geoffrey Hinton gave to Fortune magazine he explained the operations of a neural network using the example of one that is interpreting photographs, some of which happen to be those of birds.

The input would, he says, come in pixels. The first layer of neural network would detect tiny edges —dark on one side and bright on the other. This data would then pass on to the next layer, that would detect “things like corners, where two edges join at an angle”. A neuron in this layer would perhaps show a strong reaction to the angle of a bird’s beak.  The next layer might be designed to detect patterns like “a bunch of edges arranged in a circle”—flagging the bird’s head. Each deeper layer would respond to more complex concepts, detect greater abstraction, arriving finally at the concept of “bird”.

By comparing its conclusions with known photographs of birds the network would figure out if it was getting the right results. If it wasn’t messages would be sent to the layers above to retune the activation of the neurons to improve results.

This is a radical departure from the use of algorithms that are specific to the task at hand. Neural networks are completely statistical, extracting patterns from sounds and images, then comparing these to others (already labelled or analysed) to extract and extrapolate information. Where an algorithm looking to transcribe a sentence will break it down into meanings and rules of grammar, a neural network will analyse it in a manner that makes for the easiest comparison with data that it’s been trained with. This could mean that it might be working only with sound frequencies, which have little to do with either meaning or grammar. Logic and organisation are replaced here with statistical pattern analysis.

This approach worked. Proof came in October 2012 when at a workshop Fei-Fei Li, head of the Stanford AI Lab announced that two of Hinton’s students had invented an object recognition system that was twice as accurate as anything else available.

As in their biological counterparts these networks are remarkably malleable. Each layer of a neural network can be trained with different data, layers can be “peeled” and how layers interact with each other can be modified.

Best of all, nearly all neural network architecture (code), unlike most algorithms, is open source. Anyone with a little programming skill could pick up a neural network off the shelf and tweak it to an application.

 ***

T

he entry barriers to AI suddenly disappeared and the field exploded. From being niche and complex, AI became accessible and ubiquitous. While large companies like Microsoft and IBM developed their own proprietary AI architecture, this was only a minor handicap for start-ups, which could now compete with the giants.

What deep learning systems do need is lots of data to train the neural network. Data must be labelled, or annotated, so that the neural network can benchmark fresh data against it. By 2012 this was no longer a problem. The Internet was awash with data, but an even richer source of data had appeared—the smartphone—the apps for which could provide an astonishing amount of data—from what you ate to your banking habits to the music you listen to, and how frequently you travel.

Data was everywhere, and this “big data” which would have stumped most algorithmic learning systems was just what neural networks needed to extract insights far beyond human capacity. Companies like Google, Microsoft, Facebook, large banks, insurers, etc., with access to large amounts of reliable data, were uniquely placed to exploit it. Deep learning AI systems could be used in retail, banking, medicine, agriculture, everywhere. Even better could be achieved by combining task specific algorithms with neural networks.

***  

H

yperbole is perhaps the word to describe the reactions of industry titans to this new AI. Andrew Ng says, “AI is the new electricity—every major industry was transformed (by it). We now see a surprisingly clear path for this transformation.”

Tesla’s boss Elon Musk thinks AI is “more dangerous than North Korea”.

In January 2018 Google’s CEO Sundar Pitchai called it “one of the most important things that humanity is working on. It’s more profound than, I don’t know, electricity or fire.”

These extreme sentiments are echoed in India, where Gautam Shroff, chief scientist and head of research at Tata Consultancy Services (TCS) said in a phone interview that AI was the “next productivity enhancer after computing itself”.

These prophecies become more realistic against the backdrop of rapid developments in the last five years. In 2012 Google had two deep learning projects. Today it has more than 1,000 across all its major products like search, Android, maps, translation, etc. In 2016, Google’s deep learning program AlphaGo defeated the top rated Go player. Of Chinese origin, Go is considered one of the most complex board games.

According to Demis Hassabis, head of DeepMind, the division that created the program, AlphaGo taught itself largely from self-play (it played a million games against itself) and from observing professional games.

Just a year later, in 2017, DeepMind branched out to chess, pitting the program AlphaZero against Stockfish, the highest-rated chess engine. AlphaZero was not taught the game through algorithms that dissected various board positions, nor did it have pre-programmed opening and closing positions. It was just given a vast repertoire of games to analyse and four hours to “teach” itself to play the game. Compared to Deep Blue’s analysis of 200 million positions a second and Stockfish’s 70 million, AlphaZero looked at 80,000 per second—relying more on a “human-like” approach to focus on the more promising variations at any point in the game.

Companies like PayPal are now using AI-based  transaction analysis systems to prevent fraud and money laundering. In 2016 JP Morgan Chase introduced a software to review commercial loan contracts that reduced the time taken to review the 12,000 contracts that the bank handles every year from  360,000 hours to just a few seconds. When it comes to cognition and problem solving AI systems have already won the race.

There’s been rapid progress in perception, too. Speech recognition, once the frontier of AI, is almost taken for granted. Technologies like Siri, Alexa and Google Assistant are now more accurate than humans. In just one year, between 2016 and 2017, the error rate of these three dropped from 8.5 per cent to 4.9 per cent, according to a study by Stanford scientist James Landay.

Image recognition too has improved dramatically. Smartphone apps now recognise most images from birds to cars. Online social apps automatically recognise the faces of most of your friends. KFC has teamed up with Baidu, China’s Google equivalent, to offer image recognition in its restaurants that will (when you stand in front of it) not only recall your last order but also suggest things based on the kind of mood it senses you’re in. Driverless cars, once the stuff of science fiction, now roam the streets of many cities. The sophisticated vision systems used in these cars err less than once in 30 million frames (cameras record 30 frames a second) in recognising objects like a pedestrian. Till a few years ago that rate was once in 30 frames.

***

S

ubbarao Kambhampati, president of the Association for Advancement of Artificial Intelligence (AAAI), the world’s leading organisation of AI scientists, has worked in the field since the early 1990s. AI, he points out, has developed in a manner almost diametrically opposite to how a baby develops. The first intelligence the latter develops is the emotional bond with the mother, then comes social intelligence followed by perception and, much later, the ability to reason. AI, on the other hand, has mastered cognition, made progress in perception, can mirror emotional behaviour, and is taking very tentative steps towards social intelligence.

This has little to do with the difficulty of the tasks involved. Instead, it stems from a paradox pointed out by the Hungarian-British polymath Michael Polanyi: “We know more than we can tell”. For example, we know how to sort objects by sight, but would find it impossibly difficult to tell a machine how to do the same.

The corollary for Kambhampati of this paradox is “What we do well, we take for granted”. Emotions, social interactions and perception come easily to us, so we’ve only recently started exploring their deep complexity. On the other hand we find chess (or reasoning) difficult, which is why a baby learns it last—and why we’ve spent so much time studying it. Because we’ve analysed cognition intensively we’re able to “tell” computers how to emulate it.

“Dealing with each other is far more difficult than playing Chess or Go”, says Kambhampati. You get a glimpse of this in the case of people who have great artistic talent, or are autistic, excelling in areas like music or maths, but terrible at all “normal” human interactions.

*** 

I

n a cramped office in the basement of Mahajan Imaging, one of the most advanced radiology clinics in Delhi, a group of young techies pores over screens displaying X-rays.

“They’re collaborating with us to develop an AI-based X-ray analysis system,” says Vidur Mahajan, the young, Wharton-educated second-in-command at the clinic. Before this project the group had, he says, created a deep learning AI application to detect plant diseases from photos—which they sold to a large chemicals company.

The group had no medical experience, but worked with open source AI deep learning architectures that could with the appropriate training attempt radiological analysis. Mahajan Imaging’s new R&D wing—the Centre for Advanced Research in Imaging, Neuroscience and Genomics (CARING)—is working with over a dozen such groups, Indian and foreign. All are trying to create automated AI radiological imaging analysis solutions—programs that can analyse X-rays, MRIs, etc., for signs of conditions like TB and cancer, obviating the need for a radiologist and speeding up the process.

CARING provides these groups, most of them start-ups, with “10 years worth of well labelled digital imaging data,” Mahajan says. It also gives them the clinical inputs needed to train neural networks—indicators to look out for in images, false leads, other factors that aberrations can be correlated with, etc. Mahajan is optimistic about the potential of deep learning AI in India’s healthcare—making it faster, cheaper, more organised and accessible—but is also brutally honest about its current state. On their websites (like Wysa) most AI start-ups in medical imaging claim their products are in various phases of clinical testing, on the verge of deployment, but Mahajan is sceptical.

“There is not a single AI application I can use in my clinic or will be able to soon,” he says emphatically, “Not one start-up has made any money as yet, they’re all surviving on venture capital. I’m often appalled by their lack of clinical insight.”

So why is CARING working with so many? He smiles. “AI is here to stay. I want to be someone who drives that change when it comes, as opposed to following it.”

This is part of what Kambhampati, professor in the Department of Computer Science and Engineering at Arizona State University, calls “the mad rush to deployment of new technology”.

India is at the very top of this rush. A survey by consulting firm Capgemini in 2017 found that 58 per cent of companies using AI in India are using it at scale—going beyond pilot projects to deploying at commercial scale. The next largest AI adopter, Australia, has 49 per cent companies using AI at scale, while the US lags at 32 per cent.

Little of what is deployed, however, is developed by Indian companies or institutions. Most of it comes from multinationals like Accenture, Microsoft, Philips and GE Healthcare. Ironically, many of them have massive R&D centres in India, with thousands of scientists. But the annual revenue from the AI industry in India is just $180 million, according to a recent report by the Commerce Ministry’s Task Force on Artificial Intelligence. In comparison, market data company Statista pegs global revenues from AI to be $1.62 billion in 2018.

*** 

I

CICI Bank is successfully using a bot to sort customer emails on the status of transactions, etc., handle disputed ATM transactions, reducing the average resolution time from six to one hour. Nearly 40 per cent of the bank’s processes are now fully automated according to Anita Pai, senior general manager and head, operations. HDFC Bank uses a web chatbot Eva to give customers immediate information on products.

Nearly all sectors—from telecom to energy and retail—use some form of AI. Online retailers like Tata CLiQ use it to sort customer preferences and manage inventory. Ride sharing service Ola has partnered with Microsoft to predict where the demand for cabs is likely to be higher, do predictive maintenance of their fleet, etc. Hospitals like Manipal Hospital have tied up with IBM to use their AI platform Watson to help oncologists chart treatment courses for their patients.

Microsoft is venturing into new areas in collaboration with other organisations. They’ve partnered with the International Crop Research Institute for the Semi-Arid Tropics to create an “AI Sowing App” that tells farmers the best time to sow their crops. The app, currently being tested in Telangana, Maharashtra and Madhya Pradesh, looks at historic climate data spanning 30 years, and current rainfall and soil moisture levels to guide farmers.

Another app, called the Pest Risk Prediction, developed with United Phosphorous, claims to give farmers advance warning on attacks by pests like jassids, whitefly and aphids, based on crop stages and weather conditions. In the first phase 3,000 farmers in the same states receive advisories for their cotton crop. Yet another program being developed along with the Karnataka Agricultural Price Commission (KAPC), a government body, uses satellite images along with inputs on sowing area, production, weather, etc. to predict crop yields and the timing of their arrival in the market. This program, being applied to tur dal production, helps KAPC determine the price of the crop.

*** 

A

nil Bhansali, managing director of Microsoft’s R&D wing in India says their “goal is to enable data-driven farming. We are building unique solutions to solve these problems using low-cost sensors, drones, and vision and machine learning algorithms”. The company has, he says, “seen great success with our initiatives” and farmers have enjoyed the benefits. The use of AI in agriculture could transform the lives of millions of farmers in India and the world over.

Healthcare seems to be another focus area. The Microsoft Healthcare NExT initiative is aimed at “accelerating healthcare innovation through artificial intelligence and cloud computing,” Bhansali says. In India they’ve partnered with the L.V. Prasad Eye Institute in Hyderabad to come up with predictive models that look at vision impairment and eye disease in children; and with Apollo Hospitals to “predict patient risk for heart disease and assist doctors on treatment plans.”

While it’s too early to talk about the effectiveness of these deployments there have been failures. HDFC Bank’s automated email sorting system did not work. And though the collaboration between Manipal and IBM seems to be working so far, a similar high profile collaboration between the company and the MD Anderson Cancer Centre in Texas ended in failure. More often than not, AI programs have not lived up to their exaggerated marketing.

When it comes to indigenous R&D in AI, however, India is so far behind leaders China and the US that it will be almost impossible to catch up. Research is limited to some IITs and a few other institutions. Gautam Shroff of TCS, a member of the Task Force on Artificial Intelligence, says, “ If we don’t move now we’ll be 20 years behind.”

The recently released report of this task force wants a National Artificial Intelligence Mission, but has little else to offer by way of concrete suggestions.

One of the major obstacles AI development faces in India is lack of reliable data. While companies like Facebook and Google have access to endless amounts of data—a lot of it generated from India, which has the world’s largest population of Android users—this is not accessible to Indian companies.

In areas like medicine, with the exception of places like Mahajan Imaging, there is scant reliable data. “Very few hospitals like Christian Medical College are completely digitised,” says Anurag Agarwal, director of the Institute of Genomics and Integrative Biology (IGIB) and a member of the Task Force on AI. Even the All India Institute of Medical Sciences (AIIMS), probably the largest repository of medical data in the country is only partly digitised. Where there is data, “a lot of it is distorted and has errors,” he says.

The lack of a comprehensive policy on data sharing in India (it’s only now in the works) makes matters worse. Despite this, medicine seems to be the field in which AI research in India is concentrated, and to which most start-ups flock. The reason? Untapped potential, venture capital funding and a social cause.

*** 

O

ne of the start-ups Mahajan Imaging is working with is Qure.ai, a Mumbai-based company, “whose mission is to make healthcare more accessible and affordable through the power of deep learning”. Their focus has been on automated analysis of chest X-rays to detect TB. With “over a million X-rays analysed,” says Pooja Rao, head of R&D, “we’ve achieved accuracy rates higher than 90%”.

She says the TB analysis solution is being tested in collaboration with healthcare NGOs in Cameroon, Nepal, Bangladesh and India. “It is inevitable that we’ll get better with time.”

The company is now working on the analysis of head CT scans for signs of haemorrhage. In a paper co-authored with CARING they claim it successfully identifies abnormalities in more than 90 per cent of cases.

Another company, SigTuple, initially funded by Google’s venture capital arm Google Ventures (now known as GV), is using deep learning to automate pathology. Unlike other start-ups that use off-the-shelf software, SigTuple created their own architecture called Manthana. The company now offers automated solutions for blood analysis (like complete blood counts), urine and semen analysis; and eye scans.

One of the reasons for creating their own architecture according to company CEO Rohit Kumar Pandey is that it allowed them to automate the annotation, or labelling, of data required to train their neural network. So instead of manually having to tag digitised slides with (previous) diagnosis, the platform automatically scanned printed lab reports for relevant diagnosis, and tagged it to slides.

Companies like Bengaluru-based Niramai use thermographic (thermal) images to screen for breast cancer; ChironX.ai has developed a solution called ChironEye that analyses eye scans for signs of disease, like diabetic retinopathy.

At the government run IGIB, a group of researchers led by Anurag Agarwal has been correlating the molecules found in the exhaled breath of children with subtypes of asthma.

All of this work is varied and sounds very exciting, but it’s bedevilled with problems, some of which are deep.

Even though CARING collaborates with Qure.ai, Mahajan says he wouldn’t use the automated X-ray analysis system. First, the system is right only 90 per cent of the time; his centre can’t afford to go wrong on the other 10 per cent. “If a radiologist looking at the X-ray had a doubt, he’d refer it to a senior radiologist, or a few radiologists would look at it together,” he says. According to Vasantha K. Venugopal, lead imaging specialist at CARING, “the system can’t tell whether the TB is active or inactive.” A radiologist would still have to look at the image, defeating the purpose.

Mahajan says the head CT scan analysis program, the results of which were detailed in a paper he co-authored, missed a “very large bleed” in one scan. “It’s a mistake no radiologist would ever make, it’s unacceptable. Radiologists will look at this paper and the images that go with it,” he says, “and rip it apart.”

*** 

T

he problem here, which cuts straight to the foundation of neural networks, is that it’s very difficult to pinpoint exactly where the network went wrong, to understand its “failure mode”. 

One of the problems could, as agencies using predictive law-enforcement software are discovering, be with hidden biases in the training data. In the US, investigative journalism organisation ProPublica found that an AI computer program used by a US court for risk assessment wrongly flagged black people as being nearly twice as likely as white people to commit another offence. Though the company responsible never admitted it, the most likely cause was that the training data had a preponderance of black people.

But even in cases where training data is adequate and unbiased, there’s a fear that the neural network will go off on a tangent, finding irrelevant co-relations like that “between say a collar bone and the lungs in a chest X-ray,” says Vidur Mahajan. For relatively simple analysis like chest X-rays this is easily sorted by defining parameters for the neural network but it becomes more difficult as the complexity of the task increases.

The most intractable problem—known as the “black box” problem—stems from the opacity of the operation of a neural network.

***

N

eural networks are completely statistical—comparing different details and aspects of an input with labelled reference data they’ve been trained with. There’s no causality or logic involved. Even though the layers of a network can be individually programmed—in terms of what they look at and how they feed into other layers—they “notice” so much that it’s almost impossible to figure out what connections they’re drawing and how.

And since it’s impossible to figure this out it’s impossible to know why they go wrong when they do. In the most delicious irony we’re faced with the reverse of Polanyi’s Paradox— “Neural networks know more than they can tell”.

Imagine, for instance, a hypothetical neural network designed to determine whether an airplane is capable of flying. A good one would have to be trained with images of all aircraft that have ever flown and parameters like lift, weight, dimensions, materials used, etc.

With adequate training the network will learn to distinguish significant from insignificant features. But there still remains the chance that it will one day come across a case in which a designer uses a standard material for an aircraft wing, but in a dimension that tends at a particular temperature to collapse due to stress. An aircraft designer working with a materials scientist would know this, but for the AI network it would end in catastrophic failure, and there’d be no way of telling what went wrong.

That’s what happened on March 18, 2018, when an Uber self-driving car fatally knocked down a pedestrian in Tempe, Arizona. It was the first self-driving car casualty in the millions of miles these cars have driven across cities around the world, a better record than that of any group of people. But it was inexplicable.

AI technologies are, as Subbarao Kambhampati of AAAI puts, it “brittle”. It’s why he thinks it’s problematic for an app like Woebot to be interacting with a depressed American teenager. It’s why Anurag Agrawal of IGIB says that it will be “another 5-10 years before AI can (even) give surgical advice.” It’s also why companies are hesitant to let AI apps decide loan applications.

When they work neural networks do extremely well, but in the rare cases that they fail the failures can be spectacular. The scary part is that unlike failures in other machines or humans we have no way of figuring out why they happened.

The only way to solve this, Kambhampati says, is to “put the human in the loop” of an AI system—make human input integral to it. How this will be done is something that is being worked on.

Another challenge now, according to Kambhampati, is to “integrate what happened in AI pre-2012 (in reasoning, or cognition) with developments in area like perception that have happened later.”

“We have to combine learning with being told,” he says. For example, the algorithm to help a robot designed to navigate a room existed pre-2012. The perceptive technologies have been perfected after. The challenge that Kambhampati talks about would be for the robot to decide whether it’s easier to move around an obstacle or pick it up and move it aside. It doesn’t sound complex, but that’s because our previous knowledge of objects allows us to take this decision fairly easily. For the robot (and the scientists programming it) though it’s a decision lies at the cutting edge of AI research. 

Integrating perception with cognition is important, Kambhampati says, because only then will we truly able to communicate with machines. Till then we’d be limited to providing machines “one type of representation, expecting them to do something in a different representation”. Various top-down and bottom-up approaches have been tried to do this, but it remains tricky.

*** 

I

f you ignore the hype around AI and temper your expectations, you realise humans with AI are far better than humans without AI. The technology has already augmented human capacity across fields significantly, and will continue to do so.

As they’ve done in IT, banking and retail, AI programs can automate many tasks in medicine. They can be used, says Vidur Mahajan, to measure the size of organs like the prostrate gland in images and count brain lesions in multiple sclerosis—a tedious, error-prone task.

They can also be beneficial in public health. A consortium of institutions led by IGIB, AIIMS and IIT are using AI to collate and clean data from public health centres in several states. The aim, according to Agarwal, is to mine this information to formulate preventive health strategies.

The other most significant use of AI in medicine, akin to what it’s done in customer services, is likely to be in triaging. On February 15, 2018, American company Viz.ai became the first to get FDA clearance for an AI application that alerts doctors to the most critical stroke patients getting the attention first. In India, a similar application could reduce fatalities in overburdened government hospitals.

That AI is here to stay is clear from the controversy around job losses attributed to it. Most evidence is anecdotal and the forecasts are conflicting–information technology firm Gartner has predicted that AI will create 2.3 million new jobs while eliminating 1.8 million in 2020. Another study by research firm Forrester projects that by 2027 AI will displace 24.7 million jobs but create 14.9 million.

An India-specific study commissioned jointly by the Federation of Indian Chambers of Commerce and Industry (FICCI) and the National Association of Software and Services Companies (NASSCOM) forecasts that 9 per cent of India’s workforce will in a few years be deployed in jobs that don’t exist today, while 37 per cent will be in jobs that have very different skill sets.

Experts in the field also differ. Kambhampati thinks countries like the US will see a decline in jobs before things stabilise, while in India there will be increase followed by a dip.

“Throughout history,” says Microsoft’s Anil Bhansali,“the emergence of new technologies has been accompanied by dire warnings about human redundancy. More often, however, the reality is that new technologies have created more jobs than they destroyed.” “New jobs will emerge as AI changes how work is done and what people need from the world around them. What’s more, AI will create jobs that we cannot yet even imagine.”

The first wave of jobs in India has been spawned by AI’s insatiable appetite for annotated data. In a nondescript building in Metiabruz, a suburb of Kolkata that has the permanently unfinished look of poor urban areas in India, many young women and men sit outlining shapes of cars, trees and other objects in street images that flash on their screens. This data will be sent to a manufacturer of driver-less cars to help train their AI program.

iMerit, the company that employs them runs similar centres in small towns  in West Bengal, Odisha and Jharkhand. Their workforce, according to Jai Natarajan, head of technology and marketing, is 1,300. It is expected to double by the end of 2018. Most of these people are “millenials from the marginalised society,” he says. “They come from families with sporadic income in the unorganised sector.” Most have only studied up to class 12.

The company brands itself a for-profit social enterprise but says it is giving the “bottom of the pyramid” the opportunity to skill itself. Its clients, mostly in the US, include a Fortune 10 web services company and a global financial data and media company.

The need for training data, Natarajan says, has increased exponentially in the last few years. “From ‘once and done’ batches of 100,000+, we are now in the age of several training data sets of 1,000,000+ data points that have to be integrated iteratively into a machine learning model”.

Another company called Samasource employs people in data annotation in Kenya, Uganda, Haiti and India. Gurugram-based AI Touch has gone a step ahead to do data annotation in Chinese.

Kambhampati dismisses these jobs as being “low paid and temporary”. But he is convinced, “AI will lead to a great social churn, a reformulation of societies. The distinction between high caste-low caste, white collar-blue collar jobs will be replaced by that between routine and non-routine jobs.”

Routine work of all kinds–like aspects of investment banking and radiology–will get automated, but jobs that “we’ve undervalued, like that of a healthcare worker looking after the elderly—will become better paid”. 

Journalists like me, I like to believe, rest firmly in the non-routine.

On that happy note I thought I’d say bye to my penguin buddy Wysa. “I’m really happy that my article on artificial intelligence got done,” I told it. “Okay, write your thoughts down,” it said like an exasperated parent trying to calm a giddy child, “One thought at a time, in the text box as they come to you. I’ll list them down and we can organise it after…”