NVIDIA on A.I.s, Deep Learning and their place in IoT
Last month, Singapore Expo hosted the International Exhibition & Conference On the Internet of Things (IoT). And in a conference about IoT, it would be hard to imagine that a graphics company would have a speaker scheduled to give a talk. Except this is NVIDIA that we’re talking about, and the speaker is Dr. Simon See, Chief Solution Architect of NVIDIA, and he was there to talk us through a short history of artificial intelligence (A.I.), NVIDIA’s Deep Learning (DL) solutions in advancing A.I. research, and implementing all this into the realm of IoT.
For all our fears of A.I., we can be pretty reliant on them, if the usage of the Google Assistant, Apple’s Siri and Microsoft’s Cortana suggests anything. And since man fears the unknown more than anything else, maybe shedding some light on the subject will help alleviate these fears, and See did it brilliantly in his keynote speech.
Predictive / Passive A.I.
To understand the different developments and advancements in the field of A.I., perhaps the best way to begin is by understanding the different types of A.I. that mankind has built. And we start with the predictive A.I. As it is, it is likely the most common sort of A.I. the average person encounters, and does things that you may not even attribute to an A.I. sometimes. One of the things that a predictive A.I. does is image classification. To put it simply, show an A.I. a picture and it tells you what it thinks the picture is about.
The most common example that you’d use today is that of Google Search, but there other ways it is used as well. For example, Amazon AWS cloud services use NVIDIA’s GPUs to train smarter A.I. models and deliver more responsive user experiences with faster inferencing. In other words, NVIDIA GPUs are used to help the A.I. identify which matches it made were accurate. Not only does this make for up to 50 percent higher user engagement, it also means that the A.I. gets smarter, as the more people feed it photos, the more it learns and the more accurate it becomes.
And it’s tech like this that is used in most efforts to make a self-driving car. Using what is called semantic segmentation, the A.I. needs to be able to differentiate every single thing that it sees, from pedestrians and stray animals, to lampposts and other vehicles. Of course, identifying the object is the first part of the operation. The A.I. will also need to track their movements in real-time and predict future movement. All this will be repeated over hundreds of hours and thousands of kilometers to refine the learning process of the A.I.; NVIDIA’s A.I. hardware solution in such scenarios is the Drive PX2.
Beyond image classification, A.I.s also need to learn what is termed natural language processing. Because recognizing an image is not really enough for practical application, an A.I. will also need to understand not only the full context of an image, but also the words used to describe it. Even in this, A.I.s have made significant advancements to be able to fully describe an image accurately. Or in other words, to give an image its 1,000 words. Of course, there are some mixups, but we’re sure that it will be ironed out in time.
Recognizing the elements on an image is one thing, but understanding language is another. With the capabilities of an A.I. in recognizing images, it would still be unusable if a human cannot speak to it. On this front, Baidu has one of the fastest language-learning A.I.s out there. Powered by NVIDIA GPUs, Baidu uses what is called Deep Speech 2, which is the first speech recognition model that understands both English and Mandarin. And considering the number of users it has access to, the amount of learning that it goes through can be pretty tremendous. Andrew Ng, Chief Scientist at Baidu Research also says that DL “has pretty much taken over speech recognition”.
In practice, and among other applications, natural language processing takes place in one of the most important areas of human life: healthcare. A.I.s that have gone through sufficient training can look through a lot more scans than humans. With this, diagnosis can be done a lot quicker by the A.I., and it leaves human doctors to verify the findings of the scans identified by the A.I. to be anomalous.
The way all this is done? By feeding a load of public scans to train the A.I. for it to recognize the disease. One of these A.I.s has got over 700,000 records at its disposal, processes using the NVIDIA Tesla K80 GPU and the CUDA programming model.
Of course, A.I. application in medicine is not limited to identifying cancer growths - they can also be used to predict the onset of diseases, helping people prevent them rather than cure them. There is one such A.I., called Deep Patient. It analyzes electronic health records to predict the onset of up to 78 different diseases. And it doesn't do it with a single scanning template either - each prediction is customized to each individual patient so that the predictions are accurate to the individual instead of the general population. This allows the A.I.s to predict up to a whole year in advance before the disease actually manifests in a patient.
The same thing applies to other industries as well. The example NVIDIA gave was that of gas turbine combustors. Checking the old-fashioned way usually would only help identify a component that is already failing. An A.I., on the other hand, can manage to detect parts that are about to fail, and can do so up to 25 percent more accurately.