The Coder and the Dictator – Moneycontrol

Just after midnight one Tuesday in early 2018, the vice president of Venezuela commandeered the nations TV airwaves. Looking composed despite the hour, in a blue suit and red tie, he announced that the government was about to make history by becoming the first on Earth to sell its own cryptocurrency. It would be known as the Petro.

Three blocks away, in the vice presidents sprawling offices, Gabriel Jimnez was sitting blearily at an enormous glass conference table, pounding away at a laptop. Powerful air-conditioners chilled the air to a crisp. Lanky, with big black glasses set between a scruffy beard and a receding hairline, Jimnez had spent months designing and coding every detail of the Petro. Now, alongside his lead programmer, he was racing to make it operational, despite the fact that basic decisions had still not been made.

Just after the vice president signed off the air, his chief of staff burst into the office, furious. Jimnez couldnt understand something about typos on a website, an embarrassment to the nation. The chief brought in two guards, armed with military rifles, and told Jimnez and his programmer that they were forbidden to leave. If they made any attempt to communicate with the outside world, they would be on their way to El Helicoide. It was a distinctly Venezuelan symbol of terror: a futuristic mall project, with car ramps between stores, converted into a political prison and centre of torture.

Below the table, Jimnez furtively texted his wife. Although she had recently left him, he asked her to send him a hug and to tell his father that he was in trouble.

Jimnez was finally released just before sunrise. When he made it to his apartment, he burst into sobs. Before he had time to collect himself, he got a call. The president himself, Nicols Maduro, requested his presence. Jimnez walked to the presidential palace, pushing his way through the crowds outside with a sense of exhaustion and dread.

A few months earlier, the idea that Jimnez would be called before the tyrant who ruled Venezuela would have been unimaginable. Jimnez was just 27, ran a tiny startup, and had spent years protesting the dictator. Maduro had not just mismanaged his country into a financial crisis he had detained, tortured and murdered those who challenged his power.

But whatever Jimnez felt about the regime, he felt just as strongly about the potential of cryptocurrency. When the Maduro administration approached him about creating a digital coin, Jimnez saw an opportunity to change his country from within. If a national cryptocurrency was done right, Jimnez believed, he could give the government what it wanted a way to fight hyperinflation while also stealthily introducing technology that would give Venezuelans a measure of freedom from a government that dictated every detail of daily life.

His friends and family warned him that working with the regime could only end badly. The person overseeing the effort, Vice President Tareck El Aissami, had been called a drug kingpin by the US government and would soon be named to a federal Most Wanted list. Jimnez acknowledged the danger, but he talked about the Petro as a Trojan horse that would sneak in the kind of reforms that he and the opposition had been dreaming about for years.

The years 2017 and 2018 were full of drama for everyone in the crypto world, as the price of bitcoins shot up more than 1,000 percent before crashing. Billion-dollar fortunes were made and lost. But perhaps no one had as perilous a ride as Jimnez. His faith in digital currency transported him from obscurity to the centre of his countrys dark institutions of power. He found himself negotiating directly with Maduro and his top deputies, who often praised his ingenuity before escalating threats to Jimnezs life drove him into exile.

The actual goal of the project was to change the economic model of the oppressive regime, he told The Times recently. This was my mission and my gamble, in a bet that ended costing everything I had in my life: my friends, my partners, my reputation, my love, my company and my country.

Jimnez has been identified as the author of the Petro before, but he has never told his story. This account is based on hundreds of pages of confidential emails, text messages and government documents, as well as interviews with more than a dozen people who were involved with the project. Many spoke on the condition of anonymity because they still live in Venezuela, where openly criticizing the government can quickly lead to prison or death.

Jimnez was part of an educated class that was naturally drawn to the opposition. After college in Caracas, Jimnez spent a few years in the United States studying, getting married and doing what he could to oppose Chvez and his successor, Maduro. He also interned for a Republican congresswoman from Miami who regularly criticised the Venezuelan regime. When reformers won parliamentary elections in 2015, Jimnez felt compelled to return to his country to take part in the political opening.

Jimnez and his wife landed in Caracas in early 2016 and found a nation on the brink. Oil prices had plunged, sending Maduro into a money-printing frenzy. As bolvares became worthless, medicine disappeared, refugees drowned and children starved.

Jimnez was fairly insulated. He had founded a startup, The Social Us, that connected Venezuelan programmers and designers with US companies looking for cheap labour. Like many wealthier Venezuelans, Jimnez kept almost all his money in dollars, but this made transactions a headache. He had to illegally swap currency every few days, and a taxi ride would require a stack of bolvares so thick that most drivers accepted only wire transfers.

The situation rekindled Jimnezs long-running interest in cryptocurrencies. He began paying his employees in a digital coin; even with the crazy volatility of the crypto markets, it was more stable than a Venezuelan bank account, and it wasnt subject to the Maduro regimes diktats. The staff at The Social Us began touting cryptocurrency as a way for ordinary Venezuelans growing numbers of whom were buying bitcoins on the street to deal with practical problems. One project they designed was a payment terminal that bypassed government limits on spending.

Initially, the Maduro regime saw Bitcoin as a threat. The technology, after all, used a decentralised network to create and move money, and no authority was in charge. But then some members of the government noticed that this cut both ways. Cryptocurrency could also be a way for Venezuela to escape sanctions levied by the United States and international organizations.

In September 2017, an official loyal to Maduro floated the idea of a digital currency backed by Venezuelas oil reserves. This was unorthodox: One of the tenets of Bitcoin is that its value does not derive from a natural resource or government fiat, only the laws of mathematics. But the distinction faded in the face of Venezuelas desperation. The official, Carlos Vargas, read about Jimnezs crypto work in a local publication and asked for a meeting.

Soon the hulking form of Vargas arrived at the office of The Social Us. As he consumed an entire bag of potato chips, Vargas flattered the young digital workers, saying they were among the only people in Venezuela capable of creating what he had proposed. The idea was exactly what Jimnez had hoped to hear. The goal was to create a new Venezuelan currency that would move freely over an open network, like Bitcoin. The government would be unable to control or bungle it. Vargas wanted to call it the Petro Global Coin, but Jimnez suggested something simpler: the Petro.

The Social Us put together a short pitch deck for the Petro project. But Venezuela is filled with people proposing crazy schemes, and Jimnez didnt put too much stock in it. Then, in early December, when Jimnez was at a conference in Colombia, he got an urgent text. Maduro had just announced a national cryptocurrency called the Petro. Jimnez threw open his laptop and found a video of the president, in his usual workmans shirt, telling a whooping crowd, This is something momentous.

Jimnez dashed off a message to Vargas: Did they just steal our project?

Vargas replied, This is the project. They just approved it. Come back right away.

The vice president was friendly and curious, and suggested that this was Jimnezs project they were just there to learn from him. El Aissami wanted to know how many petros there would be and whether new ones could be mined like bitcoins. Jimnez thought that the officials didnt have a particularly clear idea of how cryptocurrencies worked.

After the call, Jimnez emailed his employees to be at the office for an early meeting. When everyone had gathered, he stood on a desk and said they should drop all other projects and focus on the Petro. People were free to leave, he said, but if they did this right, it was a once-in-a-lifetime chance to change Venezuela. We will liberate people from government controls, he said.

Jimnez opted to base the Petro on Ethereum, Bitcoins leading competitor, which would allow it to trade in the kind of free, publicly visible market that was otherwise forbidden in Venezuela. No one on the government side seemed to be worried about this or even aware of it.

As promised, Jimnez presented his plans for the Petro in late December, at a daylong conference at the central bank that included a handful of US crypto experts. When Vargas the newly appointed superintendent of Venezuelan cryptoassets got onstage, he seemed to have imbibed Jimnezs heretical views. We talk about the need to transform our system and move to a new economic system, Vargas said.

The real conversation, though, happened after the conference adjourned. Vargas told Jimnez and the Americans that the president himself wanted to meet.

It was night time, and a van took them through heavily armed roadblocks to the military base where the president kept his personal home, known as La Roca. It had a plainness that none of them had expected. An aging Chevy Camaro sat in the courtyard, next to a childs trampoline.

The air conditioner above the door was buzzing. The president asked the vice president if he would fix it. In his Adidas tracksuit, he stood on the couch and whacked the unit a few times. For Jimnez, there was a certain comfort in seeing the lack of luxury, given the privation in the rest of Venezuela.

Maduro was dressed casually, sitting on a couch with his wife, next to other top officials. He shook hands with everyone and made conversation in broken English, praising one American, Nick Spanos, for his appearance in a recent Bitcoin documentary that the dictator said he and his wife had just watched on Netflix.

Maduro told the group with a laugh that his announcement of the Petro had inspired cryptocurrency investors everywhere and helped push bitcoins to an all-time high of $20,000. It was unclear if he was joking, and everyone just chuckled.

When the president gave Jimnez the floor, he went over the basics of the Petro, including an initial issuance of $200 million. Then the finance minister spoke up, and for the first time, Jimnezs plans were challenged. The minister took out a manila folder with a map of the Orinoco Belt and said he wanted the Petro to be backed on an ongoing basis by certain oil reserves there, which were worth orders of magnitude more many billions of dollars.

Jimnez pushed back: It was one thing to tie the Petros initial price to oil, but if it couldnt trade freely after that at whatever price investors felt it was worth then it wouldnt be a revolutionary product. A Petro whose price always reflected oil reserves would essentially be a bond, and recent sanctions made it illegal for Americans to buy those.

The president didnt seem to follow the debate all that closely. As the group dispersed, Spanos did not have a good feeling about Jimnezs future. I thought he would become the scapegoat, he said later. I didnt think Id see this kid again.

Spanos remembers telling Jimnez before leaving Caracas: I wish I had a magic carpet to get you out of here.

As Jimnez watched Maduros televised talks, he was astonished by how much of what he had said at La Roca had gotten through to the president. Maduro mentioned Ethereum, white papers and transparency.

But the speeches also made it clear to Jimnez that he was no longer in control of the Petro. Maduro announced that the currency would, in fact, be tied to a specific block of the Orinoco Belt exactly what Jimnez had argued against. He complained to Vargas but was shot down: You cannot contradict the word of the president. Vargas told Jimnez to rewrite the Petros white paper to reflect Maduros decision and to do it quickly. He and the vice president were about to travel to Turkey and Qatar to begin selling the Petro to investors.

Things deteriorated rapidly. The presidents excitement turned the Petro into a project that everyone wanted to get in on, and in mid-January 2018 a series of meetings at the Ministry of Finance turned contentious. The departments top economic adviser wanted the Petro to have a stable value, controlled by the government, with an option to trade it in for actual oil. Jimnez managed to push back, winning an agreement that oil could be used to create a minimum value the state would promise to honour but that the price would also be allowed to fluctuate on open markets. He also made sure the Petro would exist on an open network of computers, tied to Ether, that would fundamentally limit the governments power to interfere.

Eventually, Jimnez became convinced that he would lose control of the project to the Finance Ministry. When he tried to resist sharing a digital copy of the white paper, he said, the minister told him by phone: You have to understand that this is now a project of the state. If you dont hand over the file, I wont be responsible for what happens to you.

Some of the staff at The Social Us worried that Jimnezs bull-headed desire to make the Petro happen put them all in danger. During another confrontation, Vargas had shown Jimnez blue folders containing intelligence dossiers compiled on the employees; after yet another dispute, triggered in part by the fact that the startup had not been paid anything, the vice president sent word to Jimnez that he now considered him a traitor.

It would have been reasonable at that point to assume that he was headed to prison and that his role in the Petro was over. And yet Jimnez was pulled back into the program in a shambolic series of events. The government told his team that they would need to compete to have a role in the Petros launch against a Russian group of murky origin. Jimnezs employees could find no evidence that they had any significant cryptocurrency experience; Time magazine later advanced a theory that they represented a Kremlin effort to control the Petro.

In any event, the Russians showed little interest in doing any work. Jimnez and his company were left to handle almost everything as the February 20, 2018, Petro launch date approached. That is how Jimnez found himself feverishly coding all night under armed guard and then summoned to the presidential palace early the next day.

I didnt know who my enemies were in there, he said later, recalling the event. I was the guy with no power.

After some chitchat, the president led everyone into a hall that had been converted to a Petro-themed television studio. With a crowd looking on, an emcee called to the stage the Russians and then Jimnez. He was presented with a pen and a contract. It was an agreement he had been refusing to sign for weeks that limited him to a role as a sales agent for the Petro a censure for his small acts of rebellion against the regime. On live television, Jimnez saw no way out. He scribbled his signature and gave a forced grin as photographers moved in.

Jimnez took a seat and wondered what he had just done. The president said that Venezuela had already collected $725 million from investors. He thanked Jimnez by name, as well as The Social Us. Its a company founded and run by young geniuses from Venezuela, the president said. You stay crazy.

The Petro never really got off the ground. On March 19, President Donald Trump signed an executive order barring Americans from using it. The same day, an Associated Press article about Jimnez noted that he had helped create the Petro for Maduro only a few years after interning for an anti-Maduro member of the House of Representatives. The congresswoman, Ileana Ros-Lehtinen, immediately wrote a letter asking the Treasury Department to investigate whether Venezuelan national Gabriel Jimnez meets the criteria to be sanctioned under the appropriate authorities.

In Caracas, Jimnez was barraged by criticism from the political left and right. The Social Us found it impossible to get new business. In July, a lawyer delivered a 68-page document to the National Constituent Assembly, asking that Jimnez be investigated for treason against the homeland.

Jimnez retreated into his apartment, and then, when he could no longer pay the rent, his mothers apartment. Friends say that they rarely saw him. Ultimately, his ex-wife persuaded him to leave Venezuela before the authorities finally decided to arrest him.

In April 2019, he sold his 2007 Toyota Autana and bought a ticket to the United States. When he arrived, he moved in with his father; in a completely separate chain of events, the elder Jimnez was waiting to begin serving a three-year sentence for his role in a money-laundering scheme at a Caribbean bank.

Jimnez spent his days putting together an application for asylum. I possess particularly features, as the creator of The Petro, that make me subject to persecution because the government wants to keep me quiet, he wrote.

Improbably, several countries had begun following Venezuelas lead and talking about launching their own government-sponsored digital currencies. China took the lead, and the European Central Bank said it was moving in the same direction. Venezuela relaunched the Petro a number of times, eventually coming out with a token given to pensioners that had none of the open properties from Jimnezs original design.

In October, Jimnez heard that he got his US work papers. He wept tears of joy. Then he got started on a new project; it involves using cryptocurrencies to help Venezuelans avoid the bolvar.

Jimnez still had essentially no money, but a crypto startup in the San Francisco Bay Area allowed him to work out of its offices, eat from the fridge and stay on a couch in the chief executives apartment. Recently, we met at a restaurant nearby. He pulled out a black notebook, in which he was writing letters of apology to the friends he had lost.

I always thought that I could find a solution, to be able to compensate for my mistakes, Jimnez had written to one of his best friends. I know that some apologies are not enough. I know I even deserve some pain, but believe me that life has taken care of giving them to me.

Time to show-off your poker skills and win Rs.25 lakhs with no investment. Register Now!

Read this article:
The Coder and the Dictator - Moneycontrol

Google is using machine learning to improve the quality of Duo calls – The Verge

Google has rolled out a new technology to improve audio quality in Duo calls when the service cant maintain a steady connection called WaveNetEQ. Its based on technology from Googles DeepMind division that aims to replace audio jitter with artificial noise that sounds just like human speech, generated using machine learning.

If youve ever made a call over the internet, chances are youve experienced audio jitter. It happens when packets of audio data sent as part of the call get lost along the way or otherwise arrive late or in the wrong order. Google says that 99 percent of Duo calls experience packet loss: 20 percent of these lose over 3 percent of their audio, and 10 percent lose over 8 percent. Thats a lot of audio to replace.

Every calling app has to deal with this packet loss somehow, but Google says that these packet loss concealment (PLC) processes can struggle to fill gaps of 60ms or more without sounding robotic or repetitive. WaveNetEQs solution is based on DeepMinds neural network technology, and it has been trained on data from over 100 speakers in 48 different languages.

Here are a few audio samples from Google comparing WaveNetEQ against NetEQ, a commonly used PLC technology. Heres how it sounds when its trying to replace 60ms of packet loss:

Heres a comparison when a call is experiencing packet loss of 120ms:

Theres a limit to how much audio the system can replace, though. Googles tech is designed to replace short sounds, rather than whole words. So after 120ms, it fades out and produces silence. Google says it evaluated the system to make sure it wasnt introducing any significant new sounds. Plus, all of the processing also needs to happen on-device since Google Duo calls are end-to-end encrypted by default. Once the calls real audio resumes, WaveNetEQ will seamlessly fade back to reality.

Its a neat little bit of technology that should make calls that much bit easier to understand when the internet fails them. The technology is already available for Duo calls made on Pixel 4 phones, thanks to the handsets December feature drop, and Google says its in the process of rolling it out to other unnamed handsets.

Here is the original post:
Google is using machine learning to improve the quality of Duo calls - The Verge

Machine Learning Could Make Government More Incomprehensible – The Regulatory Review

Misaligned incentives can encourage incomprehensibility.

The precept of the people, by the people, for the people demonstrates not only that citizens must choose leaders through accountability processes and institutions, but also that each citizens decision-making must be as informed as possible. This vision can only be accomplished through the availability of valid and relevant information that must be understandable to all audiences.

Over the last two decades, policy experts have hoped that novel technologies would be used to make information more meaningful, but many of these expectations are still unfulfilled. In her book with Will Walker, Incomprehensible!, Wendy Wagner demonstrates that various legal programs are built on the foundational assumption that more information is better, ignoring the imperative of usable and meaningful communication.

The design of many legal programs favors the production or reporting of undigested information, which is in turn passed along to an unequipped, disadvantaged audience. Wagner argues that although there are numerous procedural steps required for Congress to pass laws, there are no institutional controls that require a bill to be comprehensible to other members of Congress. This suggests that even today there remains an endemic, fundamental problem of unintelligibility.

The principle of governmental transparency is only fulfilled when information is relevant and understandable to a general audience. Unintelligible information or the mere release of unprocessed data does not fulfill the principle of transparency. On the contrary, it opens the doors wide for parties with technical expertise to profit from their strategic advantages over the less empowered. This concern is particularly relevant in the face of modern challenges, such as misinformation and the lack of actors that process information on behalf of citizens.

Automating government processes through machine learning would have uncertain implications in this regard, especially when the inner workings of those processes are unintelligible and might not benefit the average citizen, as Wagner argues.

Scholars have argued that machine learning can meet the laws demands for transparency and does not contravene the principles of nondelegation doctrine, due process, equal protection, and reason giving. It also can enhance efficacy, efficiency, and legitimacy in government. Principles of fair algorithmic governance, however, go beyond mere disclosure and understandability of the technical aspects of machine learning resources, such as the source code, data, objective function, parameters, and training data sets. Algorithmic governance is rooted in the very ecosystem over which those technical resources are applied and operate.

Thus, even if these technical resources are put into the open, they will introduce even more confusion if they are applied to a convoluted law that can only be understood by selected partieswith a narrow exception for machine learning that is applied to make laws more meaningful to a wider audience. Applying algorithm-based decisions to an ecosystem of unintelligible laws or regulations that favor a few knowledgeable stakeholders will compound any endemic problem, particularly if these very stakeholders further their agenda through their knowledge of machine learning. This situation would worsen the already fragile ecosystem to which Wagner refers.

The future of machine learning in government is therefore uncertain, because the technology is applied where processing help is needed, but also where it is convenient for stakeholders with great knowledge and agendas that might not be aligned with the average citizen.

According to Cary Coglianese, algorithms will likely be applied more often to assist, rather than to supplant human judgment. Indeed, the judiciary in the United States has been cautious and rather slow to utilize algorithms, mainly applying them to areas of risk assessment and dispute resolution. The majority of these tools are based on statistical approaches or conventional supervised learning logistic regression models, rather than unsupervised learning models.

Administrative agencies, on the other hand, seem to be way ahead of the judiciary. They have already employed full-fledged machine learning tools for various regulatory taskssuch as identifying cases of possible fraud or regulatory violations, forecasting the likelihood that certain chemicals are toxic, identifying people by facial-recognition when they arrive in the United States, prioritizing locations for police patrols, and more. As in the criminal justice system, none of this artificial intelligence has fully replaced human decision-making, with the exception of processes like the automation and optimization of traffic light and congestion avoidance, which has relegated humans to a supervisory control role, common of the automatic control field.

The application of machine learning to a government process is one of the last stages of a continuum in which algorithms become increasingly complex. This continuum starts with the processing of data which can offer meaningful visualizations, proceeds with the utilization of statistical approaches that can provide even more insights, and continues with the utilization of full-fledged machine learning approaches. The use of machine learning in governmental settings has not escaped controversy, particularly on the issues of bias, prejudice, and privacy that can arise from imperfect data. In addition to the fundamental issues Wagner addresses, various aspects of machine learning do not seem to be proper in early stages of this continuum, bringing a certain degree of pessimism about the application of machine learning in such an imperfect context.

My concerns are not unfounded. One example of the possible application of machine learning to an imperfect context is model legislation, also referred to as model bills. Unsuspecting lawmakers across the United States have been introducing these bills designed and written by private organizations with selfish agendas. For lawmakers, copying model legislation is an easy way to put their names on fully formed bills, while building relationships with lobbyists and other potential campaign donors. Model legislation gets copied in one state capitol after another, quietly advancing hidden agendas of powerful stakeholders. A study carried out by USA TODAY, The Arizona Republic, and The Center for Public Integrity found that more than 2,100 bills that became law in the last eight years had been almost entirely copied from model legislation.

Although the process of adopting model legislationor algorithmic objects, as I call them, because they could be re-utilizedcould be perfectly appropriate for bills with a proper purpose, the model bills passed into law often pursue the goals of powerful groups. Some of these bills introduced barriers for injured consumers to sue corporations, limited access to abortion, and restricted the rights of protesters, among others.

According to the study, model legislation disguises its authors true intent through deceptive titles and descriptions. The Asbestos Transparency Act, for example, did not help victims exposed to asbestos as its title implied; it was written by corporations who wanted to erect more obstacles for victims seeking compensation. The HOPE Act made it more difficult for people to get food stamps and was written by a conservative advocacy group. In all, these copycat bills amount tothe nations largest, unreported special-interest campaign, driving agendas in every statehouse and touching nearly every area of public policy, note two reporters involved with the Center for Public Integrity in its recent study.

Open Government Data, a technical and policy stance favoring publicly available government data which will facilitate the upcoming adoption of machine learning, is another area of concern. Very expensive initiatives and data portals in the United States have raised expectations but have failed to change agency resistance to openness or invigorate public participation. On the contrary, these initiatives have created barriers to access by favoring individuals and organizations with highly technical skills.

The problem of unintelligibility is not limited to the United States. An assessment of international government portals indicates that data-oriented technologies are not being used to make things more understandable, signaling to the myopic work of influential international organizations that have pushed for expensive technical implementation while leaving aside the needs of disadvantaged audiences in spite of the explicit warnings a decade earlier.

These are a few challenges the regulatory community must address to be ready for the eventual application of machine learning. Wagner is right to highlight these challenges, and her book pinpoints suggestions for addressing them at a fundamental level.

Martin J. Murillo is a member of the Institute of Electrical and Electronics Engineers and works as a control systems engineer, political scientist, and data scientist on the application of machine learning to government and politics.

Excerpt from:
Machine Learning Could Make Government More Incomprehensible - The Regulatory Review

Google TensorFlow Cert Suggests AI, ML Certifications on the Rise – Dice Insights

Over the past few years, many companies have embraced artificial intelligence (A.I.) and machine learning as the way of the future. Thats been good news for those technologists whove mastered the tools and concepts related to A.I. and machine learning; those with the right combination of experience and skills can easily earn six-figure salaries (with accompanying perks and benefits).

As A.I. and machine learning mature as sub-industries, its inevitable that more certifications proving technologists skills will emerge. For example, Google recently launched aTensorFlow Developer Certificate, whichjust like it says on the tinconfirms that a developer has mastered the basics of TensorFlow, the open-source library for deep learning software developed by Google.

This certificate in TensorFlow development is intended as a foundational certificate for students, developers, and data scientists who want to demonstrate practical machine learning skills through building and training of basic models using TensorFlow,read a note on the TensorFlow Blog. This level one certificate exam tests a developers foundational knowledge of integrating machine learning into tools and applications.

Those who pass the exam will receive aa certificate and a badge. In addition, those certified developers will also be invited to join ourcredential networkfor recruiters seeking entry-level TensorFlow developers, the blog posting added. This is only the beginning; as this program scales, we are eager to add certificate programs for more advanced and specialized TensorFlow practitioners.

Membership has its benefits. Sign up for a free Dice profile, add your resume, discover great career insights and set your tech career in motion. Register now

Google and TensorFlow arent the only entities in the A.I. certification arena.IBM offers an A.I. Engineering Professional Certificate, which focuses on machine learning and deep learning. Microsoft also has a number of A.I.-related certificates,including an Azure A.I. Engineer Associatecertificate. And last year, Amazon launchedAWS Certified Machine Learning.

Meanwhile, if youre interested in learning how to use TensorFlow, Udacity and Google areoffering a two-month course (just updated in February 2020) designed to help developers utilize TensorFlow to build A.I. applications that scale. Thecourse is part of Udacitys School of A.I., a cluster of free courses to help those relatively new to A.I. andmachine learninglearn the fundamentals.

As the COVID-19 pandemic forces many companies to radically adjust their products, workflows, and internal tech stacks,interest in A.I. and machine learning may accelerate; managers are certainly interested in tools and platforms that will allow them to automate work. Even before the virus emerged, Burning Glass, which collects and analyzes millions of job postings from across the country, estimated that jobs involving A.I. would grow 40 percent over the next decadea number that might only increase under the current circumstances.

Follow this link:
Google TensorFlow Cert Suggests AI, ML Certifications on the Rise - Dice Insights

AI cant predict how a childs life will turn out even with a ton of data – MIT Technology Review

Policymakers often draw on the work of social scientists to predict how specific policies might affect social outcomes such as the employment or crime rates. The idea is that if they can understand how different factors might change the trajectory of someones life, they can propose interventions to promote the best outcomes.

In recent years, though, they have increasingly relied upon machine learning, which promises to produce far more precise predictions by crunching far greater amounts of data. Such models are now used to predict the likelihood that a defendant might be arrested for a second crime, or that a kid is at risk for abuse and neglect at home. The assumption is that an algorithm fed with enough data about a given situation will make more accurate predictions than a human or a more basic statistical analysis.

Sign up for The Algorithm artificial intelligence, demystified

Now a new study published in the Proceedings of the National Academy of Sciences casts doubt on how effective this approach really is. Three sociologists at Princeton University asked hundreds of researchers to predict six life outcomes for children, parents, and households using nearly 13,000 data points on over 4,000 families. None of the researchers got even close to a reasonable level of accuracy, regardless of whether they used simple statistics or cutting-edge machine learning.

The study really highlights this idea that at the end of the day, machine-learning tools are not magic, says Alice Xiang, the head of fairness and accountability research at the nonprofit Partnership on AI.

The researchers used data from a 15-year-long sociology study called the Fragile Families and Child Wellbeing Study, led by Sara McLanahan, a professor of sociology and public affairs at Princeton and one of the lead authors of the new paper. The original study sought to understand how the lives of children born to unmarried parents might turn out over time. Families were randomly selected from children born in hospitals in large US cities during the year 2000. They were followed up for data collection when the children were 1, 3, 5, 9, and 15 years old.

McLanahan and her colleagues Matthew Salganik and Ian Lundberg then designed a challenge to crowdsource predictions on six outcomes in the final phase that they deemed sociologically important. These included the childrens grade point average at school; their level of grit, or self-reported perseverance in school; and the overall level of poverty in their household. Challenge participants from various universities were given only part of the data to train their algorithms, while the organizers held some back for final evaluations. Over the course of five months, hundreds of researchers, including computer scientists, statisticians, and computational sociologists, then submitted their best techniques for prediction.

The fact that no submission was able to achieve high accuracy on any of the outcomes confirmed that the results werent a fluke. You can't explain it away based on the failure of any particular researcher or of any particular machine-learning or AI techniques, says Salganik, a professor of sociology. The most complicated machine-learning techniques also werent much more accurate than far simpler methods.

For experts who study the use of AI in society, the results are not all that surprising. Even the most accurate risk assessment algorithms in the criminal justice system, for example, max out at 60% or 70%, says Xiang. Maybe in the abstract that sounds somewhat good, she adds, but reoffending rates can be lower than 40% anyway. That means predicting no reoffenses will already get you an accuracy rate of more than 60%.

Likewise, research has repeatedly shown that within contexts where an algorithm is assessing risk or choosing where to direct resources, simple, explainable algorithms often have close to the same prediction power as black-box techniques like deep learning. The added benefit of the black-box techniques, then, is not worth the big costs in interpretability.

The results do not necessarily mean that predictive algorithms, whether based on machine learning or not, will never be useful tools in the policy world. Some researchers point out, for example, that data collected for the purposes of sociology research is different from the data typically analyzed in policymaking.

Rashida Richardson, policy director at the AI Now institute, which studies the social impact of AI, also notes concerns in the way the prediction problem was framed. Whether a child has grit, for example, is an inherently subjective judgment that research has shown to be a racist construct for measuring success and performance, she says. The detail immediately tipped her off to thinking, Oh theres no way this is going to work.

Salganik also acknowledges the limitations of the study. But he emphasizes that it shows why policymakers should be more careful about evaluating the accuracy of algorithmic tools in a transparent way. Having a large amount of data and having complicated machine learning does not guarantee accurate prediction, he adds. Policymakers who don't have as much experience working with machine learning may have unrealistic expectations about that.

To have more stories like this delivered directly to your inbox,sign upfor our Webby-nominated AI newsletter The Algorithm. It's free.

Here is the original post:
AI cant predict how a childs life will turn out even with a ton of data - MIT Technology Review

Varian brings machine learning to proton treatment planning with Eclipse v16 – DOTmed HealthCare Business News

PALO ALTO, Calif., March 31, 2020 /PRNewswire/ -- RapidPlan PT is the first clinical application of machine learning in proton treatment planning

RT Peer Review is designed to streamline and accelerate the radiation therapy peer review process

Eclipse v16 has received CE mark and is 510(k) pending

Ad StatisticsTimes Displayed: 1011283Times Visited: 8805

Special-Pricing Available on Medical Displays, Patient Monitors, Recorders, Printers, Media, Ultrasound Machines, and Cameras.This includes Top Brands such as SONY, BARCO, NDS, NEC, LG, EDAN, EIZO, ELO, FSN, PANASONIC, MITSUBISHI, OLYMPUS, & WIDE.

Previously only available for photon-based radiotherapy treatment planning, RapidPlan is knowledge-based treatment planning software that enables clinicians to leverage knowledge and data from similar prior treatment plans to quickly develop high-quality personalized plans for patients. This knowledge-based planning software is now available for proton treatment planning with RapidPlan PT. The software also allows dose prediction with machine learning models that can be used as a decision support tool to determine which patients would be appropriate for proton or photon therapy. Varian is the first vendor in the industry to offer machine learning capability in both proton and photon treatment planning.

"With the number of operational proton treatment rooms continuing to increase, there is a need for experienced proton therapy clinicians," said Kolleen Kennedy, chief growth officer, president, Proton Solutions, Varian. "RapidPlan PT helps bridge the learning curve, allowing established centers to share their models and clinical experience. The machine learning in RapidPlan PT has the potential to reduce proton treatment plan optimization from a one to eight hour process, as reported by clinical proton centers, to less than 10 minutes, while also potentially improving plan quality."

In many radiotherapy departments, radiation therapy peer review meetings have been routinely integrated into the clinical QA process for safer healthcare delivery for the patient. Although the relevant patient information is manually retrievable from the clinical database, there is currently no efficient and effective platform to support these peer reviews. The RT Peer Review feature in Eclipse v16 is designed for the oncology community to seamlessly integrate this review process into their normal clinical workflow by automatically presenting the necessary information that is required for peer review.

About VarianAt Varian, we envision a world without fear of cancer. For more than 70 years, we have developed, built and delivered innovative cancer care technologies and solutions for our clinical partners around the globe to help them treat millions of patients each year. With an Intelligent Cancer Care approach, we are harnessing advanced technologies like artificial intelligence, machine learning and data analytics to enhance cancer treatment and expand access to care. Our 10,000 employees across 70 locations keep the patient and our clinical partners at the center of our thinking as we power new victories in cancer care. Because, for cancer patients everywhere, their fight is our fight.

SOURCE Varian

Go here to read the rest:
Varian brings machine learning to proton treatment planning with Eclipse v16 - DOTmed HealthCare Business News

The Future of Healthcare And AI Impact – Analytics India Magazine

Artificial Intelligence plays an important role in the pharmaceutical industry and the coming years there is simply no sign of the adoption of this cutting-edge technology slowing down. From making healthcare process automated to help in drug discovery, AI with machine learning can bring revolution in this industry. The key customer-oriented areas where AI is being implemented within the sector are the following:

Through natural language processing, audio and video files are transcribed from voice to text. These files shall be obtained from video-recordings from patients and customers speaking providing their opinion about a particular product or service. The dataset must be considerably large more than 300 audio-video files in order to assure accuracy. The larger the amount of datapoints, the better results that will be obtained.Within that process, an intelligent platform performs a sentiment analysis, which means the platform mines for a series of keywords or statements, as well as the demographics of the speaker (including gender and, possibly, age).Post-transcription, that data is categorized and classified, ready for analysis based on the chosen parameters.

Machine Learning uses diverse approaches to the creation of autonomous and supervised Neural Network-based speech recognition and translation systems. The two vanguard approaches in this period are Long Short-Term Memoryand CNN. The LTSM network or RNN has an 82 per cent accuracy score, while the vision-based Convolutional Neural Network scores 95 per cent accuracy.

Every Machine Learning algorithm takes a dataset as input and learns from this data. The algorithm goes through the data and identifies patterns in the data. For instance, suppose we wish to identify whose face is present in a given image, there are multiple things we can look at as a pattern:

Clearly, there is a pattern here different faces have different dimensions like the ones above. Similar faces have similar dimensions. The challenging part is to convert a particular face into numbers Machine Learning algorithms only understand numbers. This numerical representation of a face (or an element in the training set) is termed as afeature vector. A feature vector comprises of various numbers in a specific order.

As a simple example, we can map a face into a feature vector which can comprise various features such as:

Essentially, given an image, we can map out various features and convert it into a feature vector as:

So, our image is now a vector that could be represented as (23.1, 15.8, 255, 224, 189, 5.2, 4.4). Of course there could be countless other features that could be derived from the image (for instance, hair colour, facial hair, spectacles, etc). However, for the example, let us consider just these 5 simple features.

Machine Learning can help us here with 2 key elements:

What is the relationship between machine learning and optimization? On the one hand, mathematical optimization is used in machine learning during model training, when we are trying to minimize the cost of errors between our model and our data points. On the other hand, what happens when machine learning is used to solve optimization problems?

In simple terms, we can use the power of machine learning to forecast travel times between each two locations and use the genetic algorithm to find the best travel itinerary for our delivery truck. The following parameters need to be followed:

Here, you can predict who, and when an employee will terminate the service. Employee churn is expensive, and incremental improvements will give significant results. It will help us in designing better retention plans and improving employee satisfaction. This will be measured through the following attributes:

We will build a model that automatically suggests the right product prices. We are provided of the following information:

comments

More here:
The Future of Healthcare And AI Impact - Analytics India Magazine

Well-Completion System Supported by Machine Learning Maximizes Asset Value – Journal of Petroleum Technology

You have access to this full article to experience the outstanding content available to SPE members and JPT subscribers.

To ensure continued access to JPT's content, please Sign In, JOIN SPE, or Subscribe to JPT

In this paper, the authors introduce a new technology installed permanently on the well completion and addressed to real-time reservoir fluid mapping through time-lapse electromagnetic tomography during production or injection. The variations of the electromagnetic fields caused by changes of the fluid distribution are measured in a wide range of distances from the well. The data are processed and interpreted through an integrated software platform that combines 3D and 4D geophysical data inversion with a machine-learning (ML) platform. The complete paper clarifies the details of the ML work flow applied to electrical resistivity tomography (ERT) models using an example based on synthetic data.

An important question in well completions is how one may acquire data with sufficient accuracy for detecting the movements of the fluids in a wide range of distances in the space around the production well. One method that is applied in various Earth disciplines is time-lapse electrical resistivity. The operational effectiveness of ERT allows frequent acquisition of independent surveys and inversion of the data in a relatively short time. The final goal is to create dynamic models of the reservoir supporting important decisions in near-real-time regarding production and management operations. ML algorithms can support this decision-making process.

In a time-lapse ERT survey [often referred to as a direct-current (DC) time-lapse survey], electrodes are installed at fixed locations during monitoring. First, a base resistivity data set is collected. The inversion of this initial data set produces a base resistivity model to be used as a reference model. Then, one or more monitor surveys are repeated during monitoring. The same acquisition parameters applied in the base survey must be used for each monitor survey. The objective is to detect any small change in resistivity, from one survey to another, inside the investigated medium.

As a first approach, the eventual variations in resistivity can be retrieved through direct comparison between the different inverted resistivity models. A different approach is called difference inversion. Instead of inverting the base and monitor data sets separately, in difference inversion, the difference between the monitor and base data sets is inverted. In this way, all the coherent inversion artifacts may be canceled in the difference images resulting from this type of inversion.

Repeating the measurements many times (through multiple monitor surveys) in the same area and inverting the differences between consecutive data sets results in deep insight about relevant variations of physical properties linked with variations of the electric resistivity.

The Eni reservoir electromagnetic monitoring and fluid mapping system consists of an array of electrodes and coils (Fig. 1) installed along the production casing/liner. The electrodes are coupled electrically with the geological formations. A typical acquisition layout can include several hundred electrodes densely spaced (for instance, every 510m) and deployed on many wells for long distances along the liner. This type of acquisition configuration allows characterization, after data inversion, of the resistivity space between the wells with relatively high resolution and in a wide range of distances. The electrodes work alternately as sources of electric currents (Electrodes A and B in Fig. 1) and as receivers of electric potentials (Electrodes M and N). The value of the measured electric potentials depends on the resistivity distribution of the medium investigated by the electric currents. Consequently, the inversion of the measured potentials allows retrieval of a multidimensional resistivity model in the space around the electrode array. This model is complementary to the other resistivity model retrieved through ERT tomography. Finally, the resistivity models are transformed into fluid-saturation models to obtain a real-time map of fluid distribution in the reservoir.

The described system includes coils that generate and measure a controlled electromagnetic field in a wide range of frequencies.

The geoelectric method has proved to be an effective approach for mapping fluid variations, using both surface and borehole measurements, because of its high sensitivity to the electrical resistivity changes associated with the different types of fluids (fresh water, brine, hydrocarbons). In the specific test described in the complete paper, the authors simulated a time-lapse DC tomography experiment addressed to hydrocarbon reservoir monitoring during production.

A significant change in conductivity was simulated in the reservoir zone and below it because of the water table approaching four horizontal wells. A DC cross-hole acquisition survey using a borehole layout deployed in four parallel horizontal wells located at a mutual constant distance of 250 m was simulated. Each horizontal well is a constant depth of 2340 m below the surface. In each well, 15 electrodes with a constant spacing of 25 m were deployed.

The modeling grid is formed by irregular rectangular cells with size dependent on the spacing between the electrodes. The maximum expected spatial resolution of the inverted model parameter (resistivity, in this case) corresponds to the minimum half-spacing between the electrodes.

For this simulation, the authors used a PUNQ-S3 reservoir model representing a small industrial reservoir scenario of 19285 gridblocks. A South and East fault system bounds the modeled hydrocarbon field. Furthermore, an aquifer bounds the reservoir to the North and West. The porosity and saturation distributions were transformed into the corresponding resistivity distribution. Simulations were performed on the resulting resistivity model. This model consists of five levels (with a thickness of 10 m each) with variable resistivity.

The acquisition was simulated in both scenarios before and after the movement of waterthat is, corresponding with both the base and the monitor models. A mixed-dipole gradient array, with a cycle time of 1.2 s, was used, acquiring 2,145 electric potentials. This is a variant of the dipole/dipole array with all four electrodes (A, B, M, and N) usually deployed on a straight line.

The authors added 5% of random noise in the synthetic data. Consequently, because of the presence of noisy data, a robust inversion approach more suited to presence of outliers was applied.

After the simulated response was recorded in the two scenarios (base and monitor models), the difference data vector was created and inverted for retrieving the difference conductivity model (that is, the 3D model of the spatial variations of the conductivity distribution). One of the main benefits of DC tomography is the rapidity by which data can be acquired and inverted. This intrinsic methodological effectiveness allows acquisition of several surveys per day in multiple wells, permitting a quasi-real-time reservoir-monitoring approach.

Good convergence is reached after only five iterations, although the experiment started from a uniform resistivity initial model, assuming null prior knowledge.

In another test, the DC response measured in two different scenarios was studied. A single-well acquisition scheme was considered, including both a vertical and a horizontal segment. The installation of electrodes in both parts was simulated, with an average spacing of 10 m. A water table approaching the well from below was simulated, with the effect of changing the resistivity distribution significantly. The synthetic response was inverted at both stages of the water movement. After each inversion, the water table was interpreted in terms of absolute changes of resistivity.

The technology is aimed at performing real-time reservoir fluid mapping through time-lapse electric/electromagnetic tomography. To estimate the resolution capability of the approach and its theoretical range of investigation, a full sensitivity analysis was performed through 3D forward modeling and time-lapse 3D inversion of synthetic data simulated in realistic production scenarios. The approach works optimally when sources and receivers are installed in multiple wells. Time-lapse ERT tests show that significant conductivity variations caused by waterfront movements up to 100150 m from the borehole electrode layouts can be detected. Time-lapse ERT models were integrated into a complete framework aimed at analyzing the continuous information acquired at each ERT survey. Using a suite of ML algorithms, a quasi-real-time space/time prediction about the probabilistic distributions of invasion of undesired fluids into the production wells can be made.

The rest is here:
Well-Completion System Supported by Machine Learning Maximizes Asset Value - Journal of Petroleum Technology

What is Feature Engineering and Why Does It Need To Be Automated? – Datanami

(Dusit/Shutterstock)

Artificial intelligence is becoming more ubiquitous and necessary these days. From preventing fraud, real-time anomaly detection to predicting customer churn, enterprise customers are finding new applications of machine learning (ML) every day. What lies under the hood of ML, how does this technology make predictions and which secret ingredient makes the AI magic work?

In the data science community, the focus is typically on algorithm selection and model training, and indeed those are important, but the most critical piece in the AI/ML workflow is not how we select or tune algorithms but what we input to AI/ML, i.e., feature engineering.

Feature engineering is the holy grail of data science and the most critical step that determines the quality of AI/ML outcomes. Irrespective of the algorithm used, feature engineering drives model performance, governs the ability of machine learning to generate meaningful insights, and ultimately solve business problems.

Feature engineering is the process of applying domain knowledge to extract analytical representations from raw data, making it ready for machine learning. It is the first step in developing a machine learning model for prediction.

Feature engineering involves the application of business knowledge, mathematics, and statistics to transform data into a format that can be directly consumed by machine learning models. It starts from many tables spread across disparate databases that are then joined, aggregated, and combined into a single flat table using statistical transformations and/or relational operations.

(NicoElNino/Shutterstock)

For example, predicting customers likely to churn in any given quarter implies having to identify potential customers who have the highest probability of no longer doing business with the company. How do you go about making such a prediction? We make predictions about the churn rate by looking at the underlying causes. The process is based on analyzing customer behavior and then creating hypotheses. For example, customer A contacted customer support five times in the last month implying customer A has complaints and is likely to churn. In another scenario, customer As product usage might have dropped by 30% in the previous two months, again, implying that customer A has a high probability of churning. Looking at the historical behavior, extracting some hypothesis patterns, testing those hypotheses is the process of feature engineering.

Feature engineering is about extracting the business hypothesis from historical data. A business problem that involves predictions such as customer churn is a classification problem.

There are several ML algorithms that you can use, such as classical logistic regression, decision tree, support vector machine, boosting, neural network. Although all these algorithms require a single flat matrix as their inputs, raw business data is stored in disparate tables (e.g., transactional, temporal, geo-locational, etc.) with complex relationships.

(Semisatch/Shutterstock)

We may join two tables first and perform temporal aggregation on the joined table to extract temporal user behavior patterns. Practical FE is far more complicated than simple transformation exercises such as One-Hot Encoding (transform categorical values into binary indicators so that ML algorithms can utilize). To implement FE, we are writing hundreds or even thousands of SQL-like queries, performing a lot of data manipulation, as well as a multitude of statistical transformations.

In the machine learning context, if we know the historical pattern, we can create a hypothesis. Based on the hypothesis, we can predict the likely outcome like which customers are likely to churn in a given time period. And FE is all about finding the optimal combination of hypotheses.

Feature Engineering is critical because if we provide wrong hypotheses as an input, ML cannot make accurate predictions. The quality of any provided hypothesis is vital for the success of an ML model. Quality of feature is critically important from accuracy and interpretability.

Feature engineering is the most iterative, time-consuming, and resource-intensive process, involving interdisciplinary expertise. It requires technical knowledge but, more importantly, domain knowledge.

The data science team builds features by working with domain experts, testing hypotheses, building and evaluating ML models, and repeating the process until the results become acceptable for businesses. Because in-depth domain knowledge is required to generate high-quality features, feature engineering is widely considered the black-arts of experts, and not possible to automate even when a team often spends 80% of their effort on developing a high-quality feature table from raw business data.

Feature engineering automation has vast potential to change the traditional data science process. It significantly lowers skill barriers beyond ML automation alone, eliminating hundreds or even thousands of manually-crafted SQL queries, and ramps up the speed of the data science project even without a full light of domain knowledge. It also augments our data insights and delivers unknown- unknowns based on the ability to explore millions of feature hypotheses just in hours.

Recently, ML automation (a.k.a. AutoML) has received large attention. AutoML is tackling one of the critical challenges that organizations struggle with: the sheer length of the AI and ML project, which usually takes months to complete, and the incredible lack of qualified talent available to handle it.

While current AutoML products have undoubtedly made significant inroads in accelerating the AI and machine learning process, they fail to address the most significant step, the process to prepare the input of machine learning from raw business data, in other words, feature engineering.

To create a genuine shift in how modern organizations leverage AI and machine learning, the full cycle of data science development must involve automation. If the problems at the heart of data science automation are due to lack of data scientists, poor understanding of ML from business users, and difficulties in migrating to production environments, then these are the challenges that AutoML must also resolve.

AutoML 2.0, which automates the data and feature engineering, is emerging streamlining FE automation and ML automation as a single pipeline and one-stop-shop. With AutoML 2.0, the full-cycle from raw data through data and feature engineering through ML model development takes days, not months, and a team can deliver 10x more projects.

Feature engineering helps reveal the hidden patterns in the data and powers the predictive analytics based on machine learning. Algorithms need high-quality input data containing relevant business hypotheses and historical patterns and feature engineering provides this data. However, it is the most human-dependent and time-consuming part of AI/ML workflow.

AutoML 2.0, streamlines feature engineering automation and ML automation, is a new technology breakthrough to accelerate and simplify AI/ML for enterprises. It enables more people, such as BI engineers or data engineers to execute AI/ML projects and makes enterprise AI/ML more scalable and agile.

About the author: Ryohei Fujimaki, Ph.D., is the founder and CEO of dotData. Prior to founding dotData, he was the youngest research fellow ever in NEC Corporations 119-year history, the title was honored for only six individuals among 1000+ researchers. During his tenure at NEC, Ryohei was heavily involved in developing many cutting-edge data science solutions with NECs global business clients, and was instrumental in the successful delivery of several high-profile analytical solutions that are now widely used in industry. Ryohei received his Ph.D. degree from the University of Tokyo in the field of machine learning and artificial intelligence.

Related Items:

Are We Asking Too Much from Citizen Data Scientists?

NECs AutoML Spinoff Takes Off

Making ML Explainable Again

Read more from the original source:
What is Feature Engineering and Why Does It Need To Be Automated? - Datanami

A guide to healthy skepticism of artificial intelligence and coronavirus – Brookings Institution

The COVID-19 outbreak has spurred considerable news coverage about the ways artificial intelligence (AI) can combat the pandemics spread. Unfortunately, much of it has failed to be appropriately skeptical about the claims of AIs value. Like many tools, AI has a role to play, but its effect on the outbreak is probably small. While this may change in the future, technologies like data reporting, telemedicine, and conventional diagnostic tools are currently far more impactful than AI.

Still, various news articles have dramatized the role AI is playing in the pandemic by overstating what tasks it can perform, inflating its effectiveness and scale, neglecting the level of human involvement, and being careless in consideration of related risks. In fact, the COVID-19 AI-hype has been diverse enough to cover the greatest hits of exaggerated claims around AI. And so, framed around examples from the COVID-19 outbreak, here are eight considerations for a skeptics approach to AI claims.

No matter what the topic, AI is only helpful when applied judiciously by subject-matter expertspeople with long-standing experience with the problem that they are trying to solve. Despite all the talk of algorithms and big data, deciding what to predict and how to frame those predictions is frequently the most challenging aspect of applying AI. Effectively predicting a badly defined problem is worse than doing nothing at all. Likewise, it always requires subject matter expertise to know if models will continue to work in the future, be accurate on different populations, and enable meaningful interventions.

In the case of predicting the spread of COVID-19, look to the epidemiologists, who have been using statistical models to examine pandemics for a long time. Simple mathematical models of smallpox mortality date all the way back to 1766, and modern mathematical epidemiology started in the early 1900s. The field has developed extensive knowledge of its particular problems, such as how to consider community factors in the rate of disease transmission, that most computer scientists, statisticians, and machine learning engineers will not have.

There is no value in AI without subject-matter expertise.

It is certainly the case that some of the epidemiological models employ AI. However, this should not be confused for AI predicting the spread of COVID-19 on its own. In contrast to AI models that only learn patterns from historical data, epidemiologists are building statistical models that explicitly incorporate a century of scientific discovery. These approaches are very, very different. Journalists that breathlessly cover the AI that predicted coronavirus and the quants on Twitter creating their first-ever models of pandemics should take heed: There is no value in AI without subject-matter expertise.

The set of algorithms that conquered Go, a strategy board game, and Jeopardy! have accomplishing impressive feats, but they are still just (very complex) pattern recognition. To learn how to do anything, AI needs tons of prior data with known outcomes. For instance, this might be the database of historical Jeopardy! questions, as well as the correct answers. Alternatively, a comprehensive computational simulation can be used to train the model, as is the case for Go and chess. Without one of these two approaches, AI cannot do much of anything. This explains why AI alone cant predict the spread of new pandemics: There is no database of prior COVID-19 outbreaks (as there is for the flu).

So, in taking a skeptics approach to AI, it is critical to consider whether a company spent the time and money to build an extensive dataset to effectively learn the task in question. Sadly, not everyone is taking the skeptical path. VentureBeat has regurgitated claims from Baidu that AI can be used with infrared thermal imaging to see the fever that is a symptom of COVID-19. Athena Security, which sells video analysis software, has also claimed it adapted its AI system to detect fever from thermal imagery data. Vice, Fast Company, and Forbes rewarded the companys claims, which included a fake software demonstration, with free press.

To even attempt this, companies would need to collect extensive thermal imaging data from people while simultaneously taking their temperature with a conventional thermometer. In addition to attaining a sample diverse in age, gender, size, and other factors, this would also require that many of these people actually have feversthe outcome they are trying to predict. It stretches credibility that, amid a global pandemic, companies are collecting data from significant populations of fevered persons. While there are other potential ways to attain pre-existing datasets, questioning the data sources is always a meaningful way to assess the viability of an AI system.

The company Alibaba claims it can use AI on CT imagery to diagnose COVID-19, and now Bloomberg is reporting that the company is offering this diagnostic software to European countries for free. There is some appeal to the idea. Currently, COVID-19 diagnosis is done through a process called polymerase chain reaction (PCR), which requires specialized equipment. Including shipping time, it can easily take several days, whereas Alibaba says its model is much faster and is 96% accurate.

However, it is not clear that this accuracy number is trustworthy. A poorly kept secret of AI practitioners is that 96% accuracy is suspiciously high for any machine learning problem. If not carefully managed, an AI algorithm will go to extraordinary lengths to find patterns in data that are associated with the outcome it is trying to predict. However, these patterns may be totally nonsensical and only appear to work during development. In fact, an inflated accuracy number can actually be an important sign that an AI model is not going to be effective out in the world. That Alibaba claims its model works that well without caveat or self-criticism is suspicious on its face.

[A]n inflated accuracy number can actually be an important sign that an AI model is not going to be effective out in the world.

In addition, accuracy alone does not indicate enough to evaluate the quality of predictions. Imagine if 90% of the people in the training data were healthy, and the remaining 10% had COVID-19. If the model was correctly predicting all of the healthy people, a 96% accuracy could still be truebut the model would still be missing 40% of the infected people. This is why its important to also know the models sensitivity, which is the percent of correct predictions for individuals who have COVID-19 (rather than for everyone). This is especially important when one type of mistaken prediction is worse than the other, which is the case now. It is far worse to mistakenly suggest that a person with COVID-19 is not sick (which might allow them to continue infecting others) than it is to suggest a healthy person has COVID-19.

Broadly, this is a task that seems like it could be done by AI, and it might be. Emerging research suggests that there is promise in this approach, but the debate is unsettled. For now, the American College of Radiology says that the findings on chest imaging in COVID-19 are not specific, and overlap with other infections, and that it should not be used as a first-line test to diagnose COVID-19. Until stronger evidence is presented and AI models are externally validated, medical providers should not consider changing their diagnostic workflowsespecially not during a pandemic.

The circumstances in which an AI system is deployed can also have huge implications for how valuable it really is. When AI models leave development and start making real-world predictions, they nearly always degrade in performance. In evaluating CT scans, a model that can differentiate between healthy people and those with COVID-19 might start to fail when it encounters patients who are sick with the regular flu (and it is still flu season in the United States, after all). A drop of 10% accuracy or more during deployment would not be unusual.

In a recent paper about the diagnosis of malignant moles with AI, researchers noticed that their models had learned that rulers were frequently present in images of moles known to be malignant. So, of course, the model learned that images without rulers were more likely to be benign. This is a learning pattern that leads to the appearance of high accuracy during model development, but it causes a steep drop in performance during the actual application in a health-care setting. This is why independent validation is absolutely essential before using new and high-impact AI systems.

When AI models leave development and start making real-world predictions, they nearly always degrade in performance.

This should engender even more skepticism of claims that AI can be used to measure body temperature. Even if a company did invest in creating this dataset, as previously discussed, reality is far more complicated than a lab. While measuring core temperature from thermal body measurements is imperfect even in lab conditions, environmental factors make the problem much harder. The approach requires an infrared camera to get a clear and precise view of the inner face, and it is affected by humidity and the ambient temperature of the target. While it is becoming more effective, the Centers for Disease Control and Prevention still maintain that thermal imaging cannot be used on its owna second confirmatory test with an accurate thermometer is required.

In high-stakes applications of AI, it typically requires a prediction that isnt just accurate, but also one that meaningfully enables an intervention by a human. This means sufficient trust in the AI system is necessary to take action, which could mean prioritizing health-care based on the CT scans or allocating emergency funding to areas where modeling shows COVID-19 spread.

With thermal imaging for fever-detection, an intervention might imply using these systems to block entry into airports, supermarkets, pharmacies, and public spaces. But evidence shows that as many as 90% of people flagged by thermal imaging can be false positives. In an environment where febrile people know that they are supposed to stay home, this ratio could be much higher. So, while preventing people with fever (and potentially COVID-19) from enabling community transmission is a meaningful goal, there must be a willingness to establish checkpoints and a confirmatory test, or risk constraining significant chunks of the population.

This should be a constant consideration for implementing AI systems, especially those used in governance. For instance, the AI fraud-detection systems used by the IRS and the Centers for Medicare and Medicaid Services do not determine wrongdoing on their own; rather, they prioritize returns and claims for auditing by investigators. Similarly, the celebrated AI model that identifies Chicago homes with lead paint does not itself make the final call, but instead flags the residence for lead paint inspectors.

Wired ran a piece in January titled An AI Epidemiologist Sent the First Warnings of the Wuhan Virus about a warning issued on Dec. 31 by infectious disease surveillance company, BlueDot. One blog post even said the company predicted the outbreak before it happened. However, this isnt really true. There is reporting that suggests Chinese officials knew about the coronavirus from lab testing as early as Dec. 26. Further, doctors in Wuhan were spreading concerns online (despite Chinese government censorship) and the Program for Monitoring Emerging Diseases, run by human volunteers, put out a notification on Dec. 30.

That said, the approach taken by BlueDot and similar endeavors like HealthMap at Boston Childrens Hospital arent unreasonable. Both teams are a mix of data scientists and epidemiologists, and they look across health-care analyses and news articles around the world and in many languages in order to find potential new infectious disease outbreaks. This is a plausible use case for machine learning and natural language processing and is a useful tool to assist human observers. So, the hype, in this case, doesnt come from skepticism about the feasibility of the application, but rather the specific type of value it brings.

AI is unlikely to build the contextual understanding to distinguish between a new but manageable outbreak and an emerging pandemic of global proportions.

Even as these systems improve, AI is unlikely to build the contextual understanding to distinguish between a new but manageable outbreak and an emerging pandemic of global proportions. AI can hardly be blamed. Predicting rare events is just very hard, and AIs reliance on historical data does it no favors here. However, AI does offer quite a bit of value at the opposite end of the spectrumproviding minute detail.

For example, just last week, California Gov. Gavin Newsom explicitly praised BlueDots work to model the spread of the coronavirus to specific zip codes, incorporating flight-pattern data. This enables relatively precise provisioning of funding, supplies, and medical staff based on the level of exposure in each zip code. This reveals one of the great strengths of AI: its ability to quickly make individualized predictions when it would be much harder to do so individually. Of course, individualized predictions require individualized data, which can lead to unintended consequences.

AI implementations tend to have troubling second-order consequences outside of their exact purview. For instance, consolidation of market power, insecure data accumulation, and surveillance concerns are very common byproducts of AI use. In the case of AI for fighting COVID-19, the surveillance issues are pervasive. In South Korea, the neighbors of confirmed COVID-19 patients were given details of that persons travel and commute history. Taiwan, which in many ways had a proactive response to the coronavirus, used cell phone data to monitor individuals who had been assigned to stay in their homes. Israel and Italy are moving in the same direction. Of exceptional concern is the deployed social control technology in China, which nebulously uses AI to individually approve or deny access to public space.

Government action that curtails civil liberties during an emergency (and likely afterwards) is only part of the problem. The incentives that markets create can also lead to long-term undermining of privacy. At this moment, Clearview AI and Palantir are among the companies pitching mass-scale surveillance tools to the federal government. This is the same Clearview AI that scraped the web to make an enormous (and unethical) database of facesand it was doing so as a reaction to an existing demand in police departments for identifying suspects with AI-driven facial recognition. If governments and companies continue to signal that they would use invasive systems, ambitious and unscrupulous start-ups will find inventive new ways to collect more data than ever before to meet that demand.

In new approaches to using AI in high-stakes circumstances, bias should be a serious concern. Bias in AI models results in skewed estimates across different subgroups, such as women, racial minorities, or people with disabilities. In turn, this frequently leads to discriminatory outcomes, as AI models are often seen as objective and neutral.

While investigative reporting and scientific research has raised awareness about many instances of AI bias, it is important to realize that AI bias is more systemic than anecdotal. An informed AI skeptic should hold the default assumption that AI models are biased, unless proven otherwise.

An informed AI skeptic should hold the default assumption that AI models are biased, unless proven otherwise.

For example, a preprint paper suggests it is possible to use biomarkers to predict mortality risk of Wuhan COVID-19 patients. This might then be used to prioritize care for those most at riska noble goal. However, there are myriad sources of potential bias in this type of prediction. Biological associations between race, gender, age, and these biomarkers could lead to biased estimates that dont represent mortality risk. Unmeasured behavioral characteristics can lead to biases, too. It is reasonable to suspect that smoking history, more common among Chinese men and a risk factor for death by COVID-19, could bias the model into broadly overestimating male risk of death.

Especially for models involving humans, there are so many potential sources of bias that they cannot be dismissed without investigation. If an AI model has no documented and evaluated biases, it should increase a skeptics certainty that they remain hidden, unresolved, and pernicious.

While this article takes a deliberately skeptical perspective, the future impact of AI on many of these applications is bright. For instance, while diagnosis of COVID-19 with CT scans is of questionable value right now, the impact that AI is having on medical imaging is substantial. Emerging applications can evaluate the malignancy of tissue abnormalities, study skeletal structures, and reduce the need for invasive biopsies.

Other applications show great promise, though it is too soon to tell if they will meaningfully impact this pandemic. For instance, AI-designed drugs are just now starting human trials. The use of AI to summarize thousands of research papers may also quicken medical discoveries relevant to COVID-19.

AI is a widely applicable technology, but its advantages need to be hedged in a realistic understanding of its limitations. To that end, the goal of this paper is not to broadly disparage the contributions that AI can make, but instead to encourage a critical and discerning eye for the specific circumstances in which AI can be meaningful.

The Brookings Institution is a nonprofit organization devoted to independent research and policy solutions. Its mission is to conduct high-quality, independent research and, based on that research, to provide innovative, practical recommendations for policymakers and the public. The conclusions and recommendations of any Brookings publication are solely those of its author(s), and do not reflect the views of the Institution, its management, or its other scholars.

Microsoft provides support to The Brookings InstitutionsArtificial Intelligence and Emerging Technology (AIET) Initiative. The findings, interpretations, and conclusions in this report are not influenced by any donation. Brookings recognizes that the value it provides is in its absolute commitment to quality, independence, and impact. Activities supported by its donors reflect this commitment.

Read the original:
A guide to healthy skepticism of artificial intelligence and coronavirus - Brookings Institution