Artificial Intelligence is the hot topic in the tech world right now. Companies want to include it in some way in their platforms, students are lining up to train in it, and researchers are pushing its limits to see what they can apply it to.
And the possibilities are truly endless. Here's what I found.
Gadi Singer, VP of Intel¡¯s AI Products group, at AIDC 2018
ALSO READ:?Why Baby Diapers Are Related To Beer, And How Artificial Intelligence May Truly Hold The Answer
At Intel¡¯s AI DevCon 2018 in Bengaluru this past week, I got a brief glimpse into AI's transformative potential. Amidst the hype of an upcoming AI-focused chip was the excitement of dozens of developers and researchers gathered to exchange ideas, of which there was plenty on offer.
According to Amir Khosroshahi, Intel's CTO of AI division, there are close to 4 million AI developers in India. That¡¯s a huge number of people not just pushing the lengths to which AI can go, but also developing new ways to apply it.?
Take Kevin Doucette, for example. A film, TV, and video game composer, he¡¯s known for music in the? ¡®Dead Rising 4¡¯ video game and and the Oscar-nominated biopic ¡°Pel¨¦: Birth of a Legend¡±. Over the years, Doucette has worked with the likes of Robbie Williams and AR Rahman, but his newest partner is both highly intelligent and lacking in experience.?
Ella, the light-based avatar in the video, is a product of Doucette¡¯s work with Intel¡¯s AI research team. They trained the program extensively on guitar chords and responses to basic harmonies. The idea was to imitate how two musicians can sit down and jam together, playing off each others improvisations to craft a melody on the fly. So when Doucette plays a tune on an organ in front of him, Ella pipes back an electric guitar response immediately.?
Sure, things are a little wonky if you listen closely, but the achievement is still amazing. Ella is listening to the chord progression Doucette plays as a whole and then has to craft a response based on her training. However, at the same time, a different section of the AI is also using that output as input, and manipulating Ella¡¯s body to mimic the movements a guitarist playing that tune would display. All of that happens just a fraction of a second after Doucette¡¯s last note plays. Even more mind-boggling, all of this is being processed on the Intel Movidius, a USB-sized chip focused towards AI functions.
Other companies are also embracing AI, albeit in more traditional roles. Flipkart is just one of these examples. If you¡¯ve ever shopped on this e-commerce portal (and how could you not have), you¡¯ve encountered their AI at work every single time. For one thing, it¡¯s what powers their buying recommendations.
According to Mohit Kumar, the senior principal data scientists at Flipkart, the platforms draws about 13 million visits a week, and over 200 million pageviews a day. That¡¯s anywhere between 10 to 50TB of data generated by users everyday, and you¡¯d best believe none of it goes to waste.
Each user is given an anonymous data profile that holds everything from the devices they use to visit the portal, search filters they try, ratings you give, and even your delivery location (a broader regional location only, to pinpoint what area you¡¯re buying from).
Al of this data is processed everyday using machine learning, which then powers what suggestions you might be shown while browsing. For instance, if you¡¯ve searched for a watch recently, Flipkart might suggest a few even the next time when you¡¯re buying something else. There¡¯s a deeper mechanic at play here though. Apparently, these suggestions are influenced by the kind of purchases you make. If you¡¯re the kind to snatch up expensive branded items fairly regularly, Flipkart might suggest you look at a Michael Kors or an Armani Exchange. If your purchases tend to be more midrange however, and you look for deals, Flipkart might instead suggest a Titan or Casio.
It¡¯s easy to look at that and be offended at a faceless system judging you for your income, or lack of it. But the straight fact is that the system works. Each user generates at least 40 inferences through their browsing and buying habits, and the company boasts a 70 percent click rate on personalised recommendations based on this data.
It¡¯s also what lets you find products with very broad searches like ¡°budget laptops¡±, or ¡°red dresses¡±. Even now, the AI continues to learn how to better define and categorize products with labels like these, and you can see the results for yourself everyday. Not to mention that Flipkart is also using the AI to auto-moderate user reviews, as well as seller-provided titles for products.
On the other end of the spectrum is Philips, which is attempting to use AI to revolutionise the healthcare industry in India. As Philips Healthcare¡¯s senior director Ravi Ramaswamy points out, the healthcare costs in the country are rising, even as our population is slowly aging. What they want to do then, is to empower patients to be able to make informed decisions about their treatment, as well as make the entire process more accessible.
Reuters
Firstly, would be an app, like a few other companies are attempting right now. AI-powered apps could allow users to input their symptoms with photos, and even X-Ray scans, allowing for much easier diagnostics. In most simple cases, a trained AI here would be able to make a straightforward diagnosis based on simply data. For more complex cases, a doctor consulting over chat or a phone call is largely relying on experience and the accuracy of a patient.
When AI is brought into it, say perhaps where an MRI for a possible cancer patient is concerned, the neural network can act as the test analyser, either easily confirming the doctor¡¯s diagnosis or recommending another opinion. For most people, this would seem less efficient than perhaps just walking into a hospital, but it becomes especially beneficial for low-income households and patients in remote locations, where reaching a specialist like an oncologist is no easy task.?
Intel
Another plan currently in motion for Philips Healthcare is something they call Mobile Obstetrics Monitoring. Childbirth is susceptible to a great number of complications, and having no access to a doctor when you¡¯ve assumed all is well could spell disaster for both mother and child. That¡¯s why this AI employs a risk scoring system. Based on various data points like food intake, the family¡¯s medical history, pre-existing conditions and more, the neural network scores a woman according to how likely it is she may face complications during her delivery, or even before. Accordingly, it can direct mothers to seek medical help long before they become in dire need of it.
In small villages where the nearest delivery room is kilometres away, this could mean life or death for the mother and her baby. All of this, Philips believes, would allows the healthcare industry to be more productive as a whole, and charge patients based on the outcome of the treatment and not the attempts made. This would make the whole system cheaper, they hope, and would allow more patients to receive treatment as well as improve the quality of care.
Slightly a little more indirectly helping patients is another medical startup called EchoPixel. Though based in Silicon Valley, Intel had a demo stall available for us to check out at the event. What the startup¡¯s AI basically does is take a person¡¯s CT or MRI scans and virtualise them in 3D. Then, using special hardware, a doctor can pick out and study parts of a person¡¯s innards from all angles, improving their diagnostic capability.?
How does it work? Well, in the case of a brain cancer patient for example, it first requires getting an MRI of multiple slices of the patient¡¯s brain. These are typically represented as flat slices obviously. What EchoPixel then does is digitizes and recombines all these slices into a full 3D brain. Wearing special 3D glasses, and using a virtual stylus, a doctor can then literally pick out the brain from a patient¡¯s head, rotate, zoom it in. And because it¡¯s a combination of multiple slices, he or she can also take a section wherever they want to study independently. Though it¡¯s not fully ready yet, the entire thing could revolutionise surgery as we know it, allowing doctors to study a patient¡¯s insides, before actually diving inside. It promises to make surgery safer, and even diagnostics more clear-cut.?
According to Gadi Singer, VP of Intel¡¯s AI Products group, we¡¯ve come far in the last 5 years. Where we were first just dabbling in machine learning and understanding its capabilities with limited data sets, we¡¯re now pushing the boundaries of computing like never before. ¡°Most of the AI frameworks developers use today have arrived in just the last two years. ¡°It¡¯s the democratization of data science,¡± he says gleefully. But even as we speed towards faster and more powerful AI, there are pitfalls to beware of.
Bias for one thing is a major problem with AI, where a developer¡¯s one-sided or incomplete datasets can affect an AI¡¯s inferences. ¡°When an AI has bias, it reflects the bias of the people who designed it,¡± Singer says. ¡°The best way to tackle this is at the team level.¡± He is however confident that this is a mere speedbump, and not a complete wall across the road to full automation. ¡°In the long term, once we understand how biased data looks we can code it into the system. That way it can perform the checks and balances on itself.¡±