We use cookies and other similar technologies (Cookies) to enhance your experience and to provide you with relevant content and ads. By using our website, you are agreeing to the use of Cookies. You can change your settings at any time. Cookie Policy.

Deep Learning: Advances in Artificial Intelligence

Originally published January 31, 2017

Ron Powell, independent analyst and expert with the BeyeNETWORK and the Business Analytics Collaborative, interviews Diego Klabjan, professor at Northwestern University where he is director of management science for analytics. They discuss artificial intelligence and how the advances in technology are enabling deep learning.

Diego, at Northwestern one of your focuses is deep learning. What is deep learning?

Diego Klabjan: Deep learning is really about emulating human brains with a ďmodel.Ē That model would be very powerful and would be able to then create stronger recommendations and assistance into decision making.†

In essence, for the last twenty years, weíve been able to use so-called neural networks Ė essentially meaning you have 100 or so neurons, and one neuron passes the signal to another neuron. But in the last five or so years, scientists have been able to figure out how to actually model thousands of neurons. Each neuron passes the signal down to the next neuron. Then if the signal is strong enough, that neuron is going to pass it downstream to the next neuron and eventually down to the output.†

Just to make it clear, todayís model can handle ten thousand or even 100 thousand neurons. The presentation I did at Teradata Partners indicated that in human brains, we actually have 100 billion neurons. So the science is still way behind completely modeling our brains.

The best way to think about deep learning is with complex models that, for example, involve image or anything relating to image; text Ė understanding and analyzing text; and speech recognition Ė speech to text. These are very complex models, and you can think about deep learning as a model that takes a bunch of images and then it has the neurons that pass signals downstream. If you start with an image, at the end the last two neurons tell you if it is an image of a cat or a dog. That, in essence, is deep learning. The main driving force behind the success of deep learning in the last few years is our graphical processing units, or GPUs. They enable much more compilation of power. That led into the capability of analyzing or modeling 10 thousand neurons instead of just 100 neurons.†

What types of applications are in use today that encompass deep learning?†

Diego Klabjan: Most of the applications today are in the domain where there is not necessarily a decision maker involved. An example thatís widely known is autonomous cars. Without deep learning, we would not have autonomous Uber cars driving in the Bay area. †

Does that scare you a little bit Ė the autonomous cars?

Diego Klabjan: No, Iím actually looking forward to it. I find driving to be one of the most unproductive tasks in my life. So if thereís a way for me to not drive, Iím going to embrace it right away.†

A driverless car has a lot of cameras and radar. All of the images and signals they produce have to be interpreted and understood in real time. Behind all of that is deep learning. It is well known that, for example, recognizing pedestrians in real time is deep learning. Thatís the driving force behind autonomous cars. Without deep learning, autonomous cars would not be possible.

Another example is if you go to Google Translate, you get two text boxes where you can translate from one language to another. Iíve used it a few times, and it is actually fairly accurate. The reason it is accurate is because there is deep learning behind the scenes. So deep learning takes an English sentence as the input to the model. Then you have the neurons that produce, for example, a Spanish sentence. This machine translation between Google and Skype now also has the capability of real-time translation of English to Spanish or Spanish to English. That is all deep learning.

One more example is when you post a photo on Facebook. It tells you that this photo is a friend of yours or that itís your son or daughter. Behind that is deep learning. Itís able to take a picture and recognize the faces, and then conclude that this is, for example, the face of your son.

What do you see for the future use cases of deep learning?

Diego Klabjan: It is interesting that you ask that question. At Teradata Partners there was a speaker from Intercontinental Hotels Group, and she mentioned robots in hotel rooms. Similar to autonomous cars, robots would not have success without deep learning. You might have heard about Amazon using robots in their warehouses. That is all deep learning, because those robots have to recognize the images in real time and make decisions. They have to interpret the images. Thatís very similar to robots in hotel rooms that the speaker was referencing. Thatís clearly something that is potentially coming.

On the business side, customer relationship management is an application that, as far as I know, today is not yet using deep learning in production, but people are talking about it. Today you take one or two data sources and join them together to make decisions for targeted marketing or pricing, but if you have ten or more data sources, including all the customer interactions and social activities, analyzing that without deep learning is very hard. Always keep in mind that as soon as you increase complexity, then the current analytics tools are not capable of handling that complexity. Currently the solution is to take one data set at a time. But once you start integrating all of them, the complexity level increases substantially. Thatís where you then need deep learning. So, I definitely see a big opportunity for deep learning in customer relationship management. † †

In general, deep learning is very valuable for anything that is very complex. There are two examples that Iíve dealt with. One is in finance with stocks. So you can take 200 stocks that are heavily correlated Ė a very messy time series Ė thatís where deep learning actually helps predict potential stock prices. Another example is electricity markets and prices. The electricity price in Milwaukee is different than in Waukegan and Chicago. Predicting electricity prices is very difficult because there are a lot of them and they are very correlated. If there is a heat wave in Milwaukee, there will probably also be a heat wave in Chicago. There is a lot of correlation there.

Thatís essentially where I see a lot of new applications using deep learning in the future.

As far as the data thatís involved in deep learning, can you comment on it? Is it a different type of data?

Diego Klabjan: It is a lot of data Ė thatís for sure. These models, because they handle a lot of neurons, they need a lot of data. Deep learning on, for example, 100 images is not going to be successful. Weíre talking about tens of millions of images that have to be stored. I think Teradata has an opportunity there to marry deep learning and their current appliances. As I alluded earlier, one needs GPUs to actually calibrate these sophisticated, complex models. It can be done in the traditional way, and that is how it is currently being done. Essentially you put your millions and millions of images in a database, and then you fetch 1,000 of them and then the next 1,000, and so on.†

But this goes against the trend of bringing computation to the data. This is really taking the data out and bringing it to computation. It is offloading the data to GPUs. I think there is an opportunity for Teradata to explore running GPUs directly on data in their appliances.

One last question Ė I know you recently joined the Business Analytics Collaborative. Why did you decide to join?

Diego Klabjan: The masters of science analytics program that I direct at Northwestern joined the Business Analytics Collaborative because I think it is a fantastic idea to put together a community of business analytics experts. We definitely plan to leverage and benefit from this community. On the one hand, we can contribute with our own thinking and experience in producing the next generation of data scientists. I hope that the corporations that are members will learn from us. At the same time, we are always eager to learn what the business community is doing. We want to stay relevant, and we definitely plan to benefit by interacting with all the business users that are part of the community. So, in general, I think it is a fantastic idea, and we are definitely looking forward to being part of it.†

Diego, thank you for explaining the potential of deep learning.

  • Ron PowellRon Powell
    Ron is an independent analyst, consultant and editorial expert with extensive knowledge and experience in business intelligence, big data, analytics and data warehousing. Currently president of Powell Interactive Media, which specializes in consulting and podcast services, he is also Executive Producer of The World Transformed Fast Forward series. In 2004, Ron founded the BeyeNETWORK, which was acquired by Tech Target in 2010.† Prior to the founding of the BeyeNETWORK, Ron was cofounder, publisher and editorial director of DM Review (now Information Management). He maintains an expert channel and blog on the BeyeNETWORK and may be contacted by email at†rpowell@powellinteractivemedia.com.

    More articles and Ron's blog can be found in his BeyeNETWORK expert channel.

Recent articles by Ron Powell



Want to post a comment? Login or become a member today!

Be the first to comment!