What on earth is machine learning?

Machine Learning (ML) is a relatively new field and has been talked about quite a bit in the technology sphere. It’s one of the most important advancements since the invention of computers. While computers were these ubiquitous machines that could do a whole lot of things, they essentially sped up the process of executing tasks and could do so tirelessly, repeatedly and accurately. The only drawback if there was one, was that they needed to be “taught” how to execute these tasks through a set of instructions called programs or applications. So if one wanted to speed up banking processes, it needed an application to be developed for it, if one wanted a faster pace of human resource functions in a company, another application would need to be developed, and so on.

Another drawback was that these applications were designed with the idea that humans would make the decisions unless it was based on logic alone. This is where most systems became flawed. Humans are notoriously incapable of consistency, unable to prevent emotions from influencing their decisions and absolutely cannot keep pace with machines.

With ML, there is no more the need to programme a computer to understand a specific process. If an ML “engine” is simply exposed to a certain process, it can “learn” by itself using statistical models and predict outcomes based on its observations. Not only can it do something really quickly and relentlessly, it can even make better decisions than humans. This has deep implications on the way our society functions and gives birth to new possibilities.

Imagine an ML engine being implemented in the field of insurance. It has the potential to predict the outcome of a certain policy with great accuracy. This will reduce the premiums for every subscriber as there will be no need to set aside huge amounts in the event of errors in estimating the likelihood of a subscriber making a claim. It can even be employed to create customised policies for every person depending on their specific profiles, needs, environmental conditions or even their social media pages and fitness tracker data!

You may have already heard of self-driving cars that are safer than any human driver and ML engines reading medical images such as X-rays and CAT scans more accurately than most doctors trained in the fields — these are additional areas where ML engines are creating massive impact not only on the profitability of businesses, but also on the quality of life for people.

The ubiquity and pervasiveness of ML can further be understood from the fact that patents for this technology were the third-fastest growing category of all patents with a growth rate of 34% between 2013 and 2017. A Deloitte Global report predicts that the number of ML pilots and implementations doubled in 2018 over 2017 and will double again by 2020. International Data Corporation (IDC) forecasts spending on AI and ML to grow from $12 billion in 2017 to $57.6 billion by 2021.

What can UX do to help?

With machines doing the heavy lifting, the relationship between humans and machines is changing and from my vantage point as a UX designer, I wondered what role we would play in defining the evolving human-machine interface. In the distant future, we will undoubtedly work with artificial intelligence systems using voice and brain-wave based interfaces. But in the near future, there are interesting opportunities because the role of the user of an application changes to that of a trainer. The role of UX designers, consequently, will be to design applications in such ways as to make this training more efficient.

ML engines are based on statistics, which means that their accuracy is entirely dependent on the quantity of background data they can use for the “observations”. Some systems that have been around for a long time have large quantities of such background data. Others that have been implemented recently don’t have the amount of data that may allow the ML engines to make accurate determinations of outcomes. In such cases, we need to generate the data that the ML engines can use to learn to predict outcomes.

There are several interesting ways of doing this and it’s probably best illustrated with an example. Imagine a bank which provides several types of loans such as personal, automobile or business loans to its customers. Each loan application first needs to be verified for completeness, accuracy of information provided, eligibility of the applicant based on the loan type, size and collateral and a thorough examination of supporting documents.

The application verifier probably clicks the “Approve” button once s/he sees the details that have been provided by the customer and cross-references that against the type and size of loan at which point, it goes over to the next person in the process. If we imagine this person as a trainer instead, they could first indicate their confidence level (on a scale of 1 to 10) in the accuracy of the information they have verified — one for completeness, one for accuracy of information and one for sufficiency of collaterals provided for the loan. So instead of one “Approve” button as a data point, the system now gets three and that too on a scale of 1 to 10 which is more than a 300% increase in the quality of data provided to the ML engine. This means that the ML engine will be as capable of predicting the outcome of this step with 300% less data than a system that is designed to have only one “Approve” button.

Just like there are some people who are better teachers in real life, there will be better trainers for ML engines as well who will provide more accurate data. But the ML engine will be able to normalise this data if we provide feedback mechanisms at the end of the process and show it which approvals resulted in loans that were paid back successfully and which weren’t. So designing that kind of a feedback mechanism at the end will be important.

There are many other aspects of design that can be applied to make all this better, but the core principle is the same — that software systems of the future should imagine the user as a trainer and design all interactions based on that idea. Doing so will make ML-based systems much more effective and accurate much earlier than when this principle is not considered.

When should I engage UX designers for my business?

A prospective client approached us to design a financial marketplace based on processes that occur in the real world. We identified the complexities that would prove to be challenges for the smooth functioning of the marketplace. First, by going online, the lead generation funnel would just become wider. Since these leads would then need to be handled by agents of the company, that point would undoubtedly become a bottleneck, affecting performance. Secondly, approvals were required at every stage of the process, thus needing human intervention at each stage. This would be the second point where the speed would be reduced.

So when should you engage a UX designer? As early as possible, so you can not only save costs, but also increase your revenues?

The post How To Use UX To Make Machine-Learning Systems More Effective appeared first on Artificial Intelligence.

To quote Eoin Treacy in ‘How to Profit from the Peaceful Rise of the Robots’: “Don’t think of robots as replacements for humans – think of them as things that will help make us better at tackling many of the problems we face.”

Never has this been more appropriate than now in the digital media industry. The three key problems that machines can fix are costs, monetisation and ad quality.

As programmatic advertising evolves at a rapid rate, the need to deploy smart technology to continue to add value has increased. In August, the New York chapter of the IAB launched a working group to help the industry navigate the impact that artificial intelligence (AI) and machine learning (ML) will have on digital advertising.

The pressure will be on every company in the complex ad tech ecosystem to show they are not contributing to costs without adding the expected value back. As a company with our roots in the SSP side of the equation, we are no different. The truth is that SSPs will have to evolve or risk becoming irrelevant – much like any player in our industry.

Many of the productivity and financial gains for publishers will remain under the hood, and will not attract major headlines.

However, we all know that a lot of advancements will only prove fruitful with the right education. Here are some of the ways that machine learning technology can be used to put more money in the hands of publishers.

Bid throttling

The volume of ads flowing into the bidding landscape has skyrocketed thanks to header bidding – a tactic that has been a game changer for publishers, but has led to challenges on the buy side. SSPs and DSPs are receiving more traffic than ever before, so the problem, for buyers in particular, is keeping infrastructure costs down.

The answer is bid throttling, which leverages machine learning to help ensure the SSP is only sending traffic to DSPs that they are interested in, thereby increasing the efficiency in how they work together.

Here’s how it happens in practice. If we are working with 150 DSPs, and 10 billion ad impressions, it could require up to 1.5 trillion bid requests being sent to the DSPs, incurring significant infrastructure costs for them, and for us, when only the one DSP will actually win the bid.

In order to keep a cap on technology costs, we use machine learning to predict which DSP is likely to win the bid. Our algorithms analyse data on a historical basis and identify which attributes actually affect a DSP’s win rate. We use machine learning to look at trends in order to ascertain what’s being bought and what’s not.

Then we adjust our systems according to that trend data. If you can predict who is likely to bid and win on an impression, and then send an impression to those DSPs, we can generate significantly fewer bid requests. Our bid throttling efforts typically result in a reduction in QPS (queries per second), an increase in win rate, plus an increase in spend per QPS for DSPs.

This approach allows DSPs to be more efficient, which ultimately benefits publishers.

Dynamic Floors

The legacy programmatic pricing waterfall system was characterised by manually set rules, and were based on aggregate historic pricing data. This approach was up-ended by header bidding.

It’s about using platform data and machine learning to predict how advertisers will value an ad impression, and we work with our publisher partners to set a price floor in the auction that we submit to the bidders.

The publisher is able to reclaim more control over the ultimate price, thus the value of the publisher’s ad inventory is protected and they can garner more information about a buyer’s willingness to pay. We have seen a noticeable increase in monetisation for publishers – compared to the old auction systems.

Ad quality

The continued growth of programmatic trading has made it difficult for publishers to maintain quality control over the advertisers and creative that appears on their sites.

Thanks to machine learning, we can predict the quality of an ad before it appears on a publisher’s site. We scan ads as they come through our platform and use machine learning algorithms to identify an ad that is harmful (malware, offensive content, high bandwidth) and block it before it is shown on the publishers’ site.

We look at more than one hundred variables about the ad, and we also overlay known bad creatives, to feed our algorithms.

Machine learning has the ability to deduce over time whether it should block an ad, and which partner to trust. We also use algorithms to help discover if the ad being sent is something the publisher won’t like, which goes a long way in protecting advertisers’ brand safety.

In the movie 2001: A Space Odyssey, HAL 9000 said: “I am putting myself to the fullest possible use which is all, I think, that any conscious entity can ever hope to do”.

Meanwhile, we’re only beginning to scratch the surface of what machine learning can accomplish. Consciousness is still a long way off but HAL 9000’s sentiment was in the right place.

The post Why machine learning will help publishers appeared first on Artificial Intelligence.

When we think about the future, we might imagine flying cars, superior technology and the potential to cure all diseases, extend our lifespan and improve the general health of world’s population. These hopes are more or less realistic, but this is so in lare part thanks to the advancement of big data and the health technology industry. Big data specifically provides an enormous volume of information which helps healthcare providers analyze data and develop some of the most significant progressions in the HealthTech industry. To help highlight the impact that big data has had and will continue to have, let’s take a look at 4 essential ways that big data impacts the HealthTech industry.

1. It helps program machine learning technology

Big data has a significant role in programming machine learning and directly assisting in machines learning themselves to help identify disease and uncover potential treatment options.

Although early in development, machine learning technology has already made significant improvements in patient treatment. One specific example of is Enlitic Company’s lung nodule detector which can achieve positive predictive values at a rate 50% higher than a radiologist. At the moment, the company is testing clinical application for various medical conditions across CT, X-RAY and MRI scans with hopes of getting a much more thorough diagnosis in less time.

Still, successful application of machine learning cannot rely only on cramming collected amounts of Big Data into algorithms and hoping for the best result; it requires selection of large and highly targeted data sets which becomes a must-have skill for practitioners at this point.

2. It improves communication and transparency

Whether it’s keeping track of patient prescriptions or maintaining accurate health records across multiple health practices, communication and transparency between patients and doctors has, at times, been fuzzy.

Fortunately, Big data can solve this issue by automating the communication and transparency between both patients, healthcare providers and researchers. Much of this improvement has been the result of implementing Electronic Health Records (EHR), which have been adopted by nine out of every 10 US physicians. With EHRs, health providers can keep track of patient visits across multiple hospitals and providers, while patients are also able to quickly access their health history, along with doctor or nurse notes and recommendations, through mobile apps.

More interesting insights on EHRs and health apps. Did you know:

• 66% of Americans use mobile apps to manage a broad range of health issues
• 61% use apps to communicate with physicians
• 46% use apps as medical reminders
• 45% rely on them to track their symptoms
• Experts predict that by 2018, 50% of mobile device users will download about 3.4 billion health apps.

3. It promotes progress in the Telehealth Industry

Big data and sophisticated analytics are helping deliver personalized care through telemedicine, also known as remote patient monitoring. This enables clinicians and caregivers to target high-risk patients and offer them individualized treatment plans which are designed to prevent hospitalization and possible re-admission by predicting acute medical events. The more data we collect, the more precise predictions we can make. From there, we can implement better prevention programs.

As you may know, Telehealth system uses smartphones, Bluetooth, Fitbits, and similar sensors to gather data about everything from a person’s blood pressure, glucose levels, weight, physical activity and medication intake.

Research has shown that patients in Telehealth equipped ICU centers have 26% lower mortality rate and get released 20% earlier. By 2019, the  National Business Group on Health predicts that this number will leap to 97%.

4. It helps us catch fraud

The National Health Care Anti-Fraud Association estimates that healthcare fraud costs the US about $68 billion annually. Fortunately, big data analytics play a significant role in fraud detection and prevention. In fact, the Centers for Medicare and Medicaid Services caught and stopped more than $210.7 million in healthcare fraud in just one year using predictive analytics.

Upon finding such major success, United Healthcare decided to transition to predictive modeling environment on a Hadoop big data platform. This development helped them identify inaccurate claims and generate a 2200% return on their big data/advanced technology.

The ability to store information and go back into a patient’s history and analyze large unstructured datasets of previous claims is the key to identifying health care fraud. Machine learning algorithms are used to detecingt patterns and anomalies such as patients who are receiving healthcare services from different hospitals in few different locations simultaneously, hospital’s short-term overuse of services, or identical prescriptions filled in multiple locations.

The post 4 REASONS WHY ADVANCING BIG DATA HELPS THE HEALTHTECH INDUSTRY appeared first on Artificial Intelligence.