March 16, 2018
4 deep learning breakthroughs business leaders should understand
It's a given that artificial intelligence will change many things in our world in 2018. But with new developments arising at a rapid pace, how can business leaders keep up with the latest AI to improve their performance?
Perhaps the best place for executives to start is gaining an understanding of deep learning. As one of the most exciting and powerful branches of AI, deep learning has led to important breakthroughs that expand the possibilities of applying AI to business problems.
First, let me provide a quick intro to the technology. Deep learning is a type of machine learning. It's a subfield of AI that deals with how computers learn as opposed to focusing on how we explicitly program them. In deep learning, researchers place concepts into a hierarchy. At each layer, a machine learns a concept and passes it to the next layer, which in turn uses it to build a more sophisticated concept. The more layers these models have (or the "deeper" they are), the more concepts they can learn, putting them at the vanguard of AI.
If that all sounds a bit complicated, don't worry -- we'll dive into concrete examples below. Here are the top four deep learning breakthroughs business leaders should be aware of, arranged from the most immediately applicable to the most cutting edge.
1. Image understanding
We can train deep learning algorithms to identify objects in an image. As of 2015, these algorithms (called convolutional neural networks) can achieve better image classification results than human beings.
Above: Samples of predictions a Convolutional Neural Network architecture made about images.
Image Credit: ImageNet dataset
So how have business leaders applied these powerful algorithms to date? One application we're all familiar with is Google Image Search. By understanding what's contained in images, Google serves up appropriate responses to search queries.
Another example is self-driving cars, which identify and respond to what they "see," enabling an entirely new industry. Deep learning models have used detailed image analysis in health care to greatly improve disease diagnoses, including diabetic retinopathy and some cancers.
As you can see, companies and researchers have applied image understanding in drastically different ways to overcome various challenges. Thinking about the kind of image data your business possesses or the ways image understanding can aid your operations could help you come up with the next great product or service based on this type of deep learning.
2. Sequence prediction
Another breakthrough of deep learning is the ability to understand sequential data, like text (a sequence of characters) or a set of observations over time. Neural network architectures built for these purposes are called recurrent neural nets.
In this scenario, a researcher would train the neural networks to look at huge amounts of past sequences, learn their patterns, and generate future sequences that follow those patterns.
We've applied sequence prediction in several domains. One early experiment showed that, by representing handwriting as a series of points with X and Y coordinates, the neural network could learn to produce new handwriting that looked real.
Above: Written by a neural net. Can we still call it handwriting?
Image Credit: University of Toronto Computer Science Department
In the field of time series prediction, here's one example that may have already improved your commute. Uber found ways to predict user demand by modeling the number of rides its customers take over time as a sequence. Now you know what algorithms to thank (or curse) when you look up how many drivers there are in your area.
Sequence prediction has proven itself with a number of different applications in business. It's well worth investigating how you can apply it to yours.
3. Language translation
Machine translation has long been a dream of AI researchers. Deep learning brought that dream much closer to reality with sequence-to-sequence architecture, which uses recurrent neural networks under the hood.
As you can see from the chart below, this architecture blows other translation techniques out of the water, with the exception (so far) of human translation:
Above: Google Sequence-to-Sequence based model performance.
Image Credit: Google Research Blog
The goal of sequence-to-sequence is to optimize for language translation. Researchers discovered the technology in 2014 and have continued to improve upon it each year. The technology now powers Google Translate and Apple's Siri. Startups are also working on using sequence-to-sequence for chatbots. This area has significant promise, but so far seems to work best when we train it on narrowly defined domains, such as customer service for an app.
As these models improve, you'll no doubt want to keep a close eye on how they could drive innovation in your own field.
4. Generative models
Our last huge breakthrough achieved with deep learning is the creation of models that generate complex data, like images that look like faces but are not actual faces. This is possible due to architectures called generative adversarial networks, which use convolutional neural nets under the hood.
Above: Images generated by a generative adversarial network from a dataset of faces of celebrities.
Image Credit: Nvidia
Generative models are perhaps the most intriguing of all four deep learning breakthroughs, though as of now, their applications in business are limited.
One early use of this deep learning breakthrough has been to aid image classification models. These models can learn to understand objects in images much more efficiently if researchers train them to distinguish real images from fake ones that a generative adversarial network generates.
As data scientists refine the uses of this breakthrough, you'll want to take note of how companies use generative models in new and exciting ways so you can begin applying their power to your own business challenges.
A final note
Each of the breakthroughs above has many open source implementations. That means you can almost always download a pre-trained model and apply it to your data. For example, you can purchase pre-trained image classifiers that allow you to feed your data through to classify new images. In this case, because the company that sold you the product has done much of the work for you, you don't need to develop the deep learning to take advantage of these cutting-edge techniques. Rather, you just need to do the development work to get models others have created to work for your problem.
Now that you have a better understanding of the capabilities of deep learning models, you're a bit closer to joining companies like Uber and Google in actually using them. Remember that the next generation of business applications of deep learning is still to come. For example, when Apple introduced the iPhone, nobody was thinking about using it for ride-sharing. Now is the time to discover new ways to apply these techniques to your own data.
Seth Weidman is a senior data scientist at Metis, a company that provides full-time immersive bootcamps, evening part-time professional development courses, online learning, and corporate programs to accelerate the careers of data scientists.