The term "artificial intelligence" (AI) has been around since the 1950s, but there has long existed a yawning gap between what people thought AI to be and what was actually possible. Since the 1960s, much of what has been considered to be AI has been a form of machine learning. However, despite the leaps and bounds in technology since the term was first introduced, the accomplishments of machine learning have often fallen short of public expectations (think the robots in The Jetsons).
Fortunately, after a half-century of research, that gap between our expectations and reality is finally closing, thanks to deep learning, a more advanced type of machine learning that's capable of generating human-like insight. Here, I will give an introduction to deep learning and explore the potential of this groundbreaking new field.
Before we dive into deep learning, we need to first define what machine learning is. Machine learning basically involves giving data to an algorithm and then "training" it to perform a certain task, such as giving out movie recommendations based on a user's past watching history. In other words, with machine learning, you plug the data in and get an answer in return. This is distinct from the older approach of having a programmer write a program to generate movie recommendations. Machine learning uses the data to write the program.
So, what is deep learning? Deep learning takes machine learning to the next level through the use of artificial neural networks. These networks are essentially layers of artificial "neurons" in a rough approximation of the way that human brains work. These neural networks operate by taking an input - say, an image of a cat - and passing it through layers and layers of neurons that each perform simple computations until ultimately the last layer of neurons output the name of the object in the image: "cat."
Until recently, neural networks were considered a fairly unsuccessful branch of artificial intelligence. All of that changed when pioneers like Geoffrey Hinton and Andrew Ng managed to create amazingly good image recognition systems built on massive neural networks that had been trained with large amounts of internet data.
Neural networks are explicitly engineered to be able to recognize complex patterns. Given enormous amounts of historical data, they are able to recognize patterns that allow them to successfully predict answers or outcomes. In today's digital era, with mobile phones and IP-addressable devices in every home, the amount of data being anonymously created by consumers about their behaviors online and on the street is enabling these neural networks to better predict people's actions. These actions include predicting how human drivers will react so a self-driving car can react quickly and safely, predicting where crime is likely to occur next, how to diagnose patients or even identify how someone makes a decision to buy a new TV.
Unlike other machine learning algorithms, neural networks are also designed to be able to learn from their mistakes. This is another reason many deep-learning scientists believe we are finally beginning to develop machines that come close to what most people would consider artificial intelligence. Just like humans, machines can now make a prediction, act on that prediction and learn from that action, whether right or wrong. Like humans, this allows them to become more accurate over time as more decisions and feedback is collected. For the first time ever, neural networks are enabling human-like insight and learning in computers.
The most high-profile example of a neural network has to be DeepMind's AlphaGo
, a neural network that was trained to be able to beat some of the best Go players in the world. Recently, DeepMind made the news again with another version of AlphaGo named AlphaGo Zero. Unlike previous iterations, AlphaGo Zero was not given any human data and instead familiarized itself with the game by simply trying out moves to see if they worked. In the words of a TechCrunch article, AlphaGo Zero "managed to rediscover over 3,000 years of human knowledge around the game in 72 hours" and went on to best the AlphaGo that had beaten Go grandmaster Lee Sedol in a mere three days. Astonishingly, DeepMind never even found the limit of AlphaGo Zeros' intelligence. As the company's CEO Demis Hassabis put it, "We needed the computers for something else."
Deep learning and neural networks are not just for creating a machine that's good at a board game; they're also what allows machines to recognize certain images, root out offensive content online and understand what we say. For example, if you use a voice assistant, whether it's Siri, Alexa, Google Home or Microsoft Cortana, you're relying on deep learning to understand your voice and interpret what you want. If you post a picture on Facebook, Facebook will use image recognition software to identify the people in the photo, relying, again, heavily on neural networks
to make the identification.
Ultimately, deep learning has pushed machine learning across a threshold. Whereas machine learning had some success in automating repetitive tasks or data analytics, it is now bringing the future to life in the form of computers that can see, hear and play all types of games. At our company, we pride ourselves on being at the forefront of that change by using advanced artificial neural networks to transform the way brands play the game of marketing.