When you asked Siri to look for the nearest restaurant, did you ever wonder how it actually worked? How could it recognize your speech and fulfill the task? The answer lies in ‘deep learning,’ which multinational companies such as Google and Microsoft are heavily investing in.
In an exclusive interview with The Smart Manager, Philip M Parker explores the nuances of this technology and how it is increasingly disrupting businesses.
How is deep learning technology going to affect businesses worldwide?
‘Deep learning’ is a layer that goes on top of other big data applications. If there
is a big data warehouse containing petabytes of data comprising images, text, and voice files among other things, deep learning is a layer that goes on top of that which allows you to extract knowledge or create new knowledge; and that is the point of
deep learning. This technology is not new—it has been in practice since the early 1970s, with the advent of neural networks.
Neural networks are algorithms that use back propagation, which was developed in the mid-1970s to make inferences from data and recognize patterns. Its earliest applications were in particle physics wherein people had to process and look at images of particles colliding with each other. Neural networks were developed to look at these images and decide whether real physical events occurred. Rather than asking doctoral students to look at these images, those at Stanford Linear Accelerator Center were using algorithms for image pattern recognition.
Most of the early work in this area was largely in the field of science, computer programming, physics, etc. Its application in business is a relatively new phenomenon. We are at the beginning of the life cycle of this concept and there are only a few companies that are actively engaged in this. Many are engaged in big data but they do not have the expertise to use Artificial Intelligence (AI) as a layer on top of it, and deep learning as another layer.
can you elaborate on the three different layers?
There are three layers—a layer of big data which comprises massive data warehouses, a layer of deep learning, and a layer on top of it—authoring layer—wherein after learning something, something original is authored. I am particularly working on the authoring layer. This concept was probably pioneered here at INSEAD, and also at MIT by Professor John Little who noticed that many were getting scanner data from optical scanners but they simply did not have the time to analyze it. Algorithms were used to detect events in the data and then a memo was sent to marketing managers, letting them know of any ‘news’ in it. It literally wrote news articles for managers, and right now we are witnessing the application of the authoring layers in a number of domains. We cannot look at deep learning in isolation, since without data we do not have deep learning, and without the authoring layer you do not get much value out of deep learning.
It was in the 1980s and the early 1990s that we witnessed the first application of all the three layers in the field of business. Deep learning is a collection of different algorithms that people use to make data representations, and to make them automatically without involving labor. Earlier, people had to physically tag photographs but with deep learning today, we have computer programs that mimic human behavior and learn as they recognize more and more patterns.
The problems that businesses face today are too vast to be solved through labor. The most effective way is to use big data with deep learning and automated authoring.
Automated authoring is undoubtedly growing rapidly. Right now, weather forecasts are available in 120 languages because of the application of the authoring layer. Most of the world’s languages did not have weather forecasts at the rural level or in the local dialect, until we created it. It has many application areas—whatever you think a human being could potentially author, so could a computer algorithm which has been trained with the deep learning algorithm. My website (totopoetry.com) uses algorithms to write poetry as well as to edit and fine-tune them. It first produced 4.5 million poems which got edited by another computer program to 1.4 million, based on quality. This algorithmic authoring has applications in many domains such as medical care (wherein a patient can type in symptoms and get a diagnosis done), crop care, and livestock care. I received a patent for this technology from the US Patent Office back in 2000. It is, however, new and is not yet prevalent across organizations worldwide. I would say less than 1/10 of one per cent of companies are using this. However, the firm that I started—the Icon Group—has published over one million books written fully by computer programs and these are distributed on Amazon.com.
This reliance on big data and related technologies is a natural progression for companies that possess the skills to exploit data and/or the requirement to do so. So it is natural for Amazon to recommend the next movie or an ecommerce player to recommend more products—this is what salespersons have been doing in stores for a hundred years. They notice what you buy and then make a recommendation. Recommendation engines often rely on a layer of deep learning. The next layer on top of that will be the authoring layer, which would actually have an artificial 3D salesperson helping and guiding users in their shopping experience.
how is deep learning different from AI?
We need to see AI as a collection of various programs. It is anything that can mimic an intelligent being. A pocket calculator that will solve basic mathematical problems is also a form of AI. It is faster than a human being and might be more accurate too. It is programmed to give answers to questions. However, deep learning does not belong to this category of AI. It is an application area of AI within which there are other applications people are working on, such as image, voice, and facial recognition; recommendation engines; tracking criminals and criminology; and writing new books or authoring original content. Most of the languages in the world do not have recommendations on how a farmer should grow crops. I am working on a project where we used an authoring engine on top of the deep learning engine to come up with [information on] optimal crops, weather forecasts, etc. These are the application areas within deep learning.
Deep learning is about representing the existing data, finding a pattern, and then extrapolating from that pattern. Finding a pattern in data is a simple thing to do. Natural language parsing is a part of deep learning. Take a sentence like ‘George Washington was the first President of the United States’; the word ‘was’ is a magic parsing word in deep learning. That one statement can answer two questions—‘who was George Washington?’ and ‘who is the first President of the United States?’ Deep learning algorithms run through billions of sentences that have ever been published in science and technology and extracts knowledge, therefore being able to answer questions.
The interesting part is whether it can author something new since it knows everything. For example, some day there could be a cure for malaria and there will be a sentence—the cure for malaria is ‘something’. Deep learning, on top of authoring, allows it to actually write— ‘the most likely cure for malaria will be.’ So that is the next frontier. Companies such as IBM and Microsoft are working in this area where we not just recognize photos but also say, “now that I know what the photographs are, I can tell where a plant disease is likely to spread the fastest.” It is about coming up with original insights and conclusions that human beings simply would not have the time to analyze. In addition, deep learning algorithms have been used for simple things such as playing a game of chess. It learns the opponent’s weaknesses and then comes up with optimal moves, similar to rules-based algorithms which get better when trained more.
which are the industries where deep learning will have a significant impact, and how?
Deep learning technology will have the most impact in image and voice recognition. It will also be used in areas such as criminology, and in the pharmaceutical industry for discovering new compounds and drugs. It will also be instrumental in developing next-generation search engines. Today, search engines use the legacy technology of spidering the whole internet and discovering the best information on something you type. However, with data growing so fast there will not be enough hard disc space to store it. So deep learning will have to replace the current search engine technology.
I am working on a project (totoGEO) funded by Bill and Melinda Gates Foundation that is utilizing deep learning techniques to bridge the content divide, and authoring new content and educational materials in underserved languages. It involves natural language processing that goes through billions of text, learn from it, and then reproduce it in all of the world’s languages. Right now, there is a major content divide between people who generate content and knowledge, and those who actually might need it the most. A bookstore in the US has a big self-help section whereas if you go to a French bookstore, there is a smaller self-help section because there are less number of people who speak French and hence it is not economical for the publisher to publish books in French. As a language gets forward, such as the 19 languages of India, for example, the less is published within those languages. If you type a word like ‘molecule’ or ‘macro-molecule’ in Hindi, there will not be any results on the internet.
Deep learning can also help in translation. In businesses, it is being used for micro-segmentation and micro-target marketing to learn more about the micro-needs of customers and their behavior, and transfer that knowledge to decision-makers so that they can shape better business strategies instead of making gut-based decisions. A retail chain, for example, has thousands of products but no one can be an expert on all. But they can leave the decision-making to the algorithms and machines. I believe that eventually, marketing managers will be replaced by computer algorithms. Even journalists will be. A US-based company, Narrative Science, writes more news articles than all journalists in the country. Many professions will be affected in the next ten years.
can you discuss the challenges one might face while adopting deep learning technologies?
Almost 98% of senior managers have never heard about deep learning. A few IT and technology managers have read about it in magazines or on Wikipedia, but are not actively engaged in it. Because right now, they are getting the first layer which is the big data layer settled. In the second layer, they might be doing some simplistic report-writing but nothing beyond that. The third layer only has a handful of companies involved in it. One of the barriers is lack of knowledge, another is that even if they have heard it before, the IT teams might not be well trained for it. This entails a combination of mathematics, computer programming, and in some cases, special skills based on the need of the project. For a sound project, we need people with knowledge of three disciplines— mathematics, computer science, and acoustics engineering—and this makes it difficult to find people qualified for it.
People feel its impact but they do not know the origin of the impact. They will hear the weather report but will not know the technology used to produce it. They might see a video being recommended on a website but are unaware of the fact that deep learning was used to come up with the recommendation. The interesting part is that deep learning has already made its impact on most people but they have not realized that it is impacting them.