In the previous part of this series which can be found here we’ve only just begun to scratch the surface of AI as a whole. Most of that post was dedicated to Machine Learning but that is only one sub category. If we peel back another layer we discover what is called “Deep Learning”
AN OVERVIEW OF DEEP LEARNING
Deep Learning is a sub field of Machine Learning dealing with specific algorithms that resemble the structure and function of the brain. This has also been referred to as “Artificial Neural Networks”. Now the term “Deep Learning” has been used to describe a wide array of actions and algorithms in the AI world and the exact definition has became rather fuzzy over the years. In this post we will attempt to cut through the hype and the cryptic nature of this sub field to shed some light on what this actually is.
The main idea behind this technology has been around for decades now. It entails simulating the neocortex’s large array of neurons in an artificial neural network. The neocortex refers to a large area of the human brain (roughly 80 percent) where most thinking occurs. A piece of software able to mimic the connectivity and the action potential of this brain region would be able to learn and find patterns across a variety of different kinds of data.
This has been attempted many times before but up until recently computers simply were not fast enough to handle the mathematical formulas required to explore this approach. The potential to model layers upon layers of neurons is now upon us and great strides have been made as a result. Google in particular has used deep learning to make extraordinary advances in speech and image recognition.
DEEP LEARNING IS A NEW NAME FOR NEURAL NETWORKS
Deep Learning is largely a revival of a 70 year old approach to artificial intelligence which was known as “neural networks”. This was discovered in 1969 but didn’t really get off the ground until the 1980’s. They were a means of performing machine learning by having computers learn to perform tasks by analyzing training samples. Even back then neural networks were based on the human brain which contained thousands or even millions of simple processing nodes that were densely interconnected.
Most of today’s neural nets are organized into layers of nodes. Each of these nodes will typically assign a number as its “weight”. When the network is active, the node receives a different data item that represents a number over each of its connections and multiplies it by the associated weight. It then adds the resulting products together which records it as a final value. If that number is below a certain threshold value the node won’t pass whatever data its holding to the next layer. It that number exceeds the threshold value, the node will “fire”, which sends the final value forward through all its outgoing connections. Each node could be connected to many different branches and layers of the neural network as a whole so that when each one fires, the data moves on in an organized and efficient way.
On paper the system as a whole works extraordinarily well, and this usually does translate to real world success, provided the neural network can be trained properly and provided with a large enough data set. More recently specific algorithms have been written to train neural networks to learn on their own with amazing efficiency. This was in large part why the AI called “Alpha Zero” was able to become the most powerful chess playing entity in the world with only hours of training. This AI was covered in a previous post.
The above graphic displays an overview of the basic principles governing a neural network.
REAL WORLD APPLICATIONS
If you combine the right algorithms with the right models then deep learning or deep neural networks can set the stage for some serious real world applications that can tackle some of the most important problems we face on a daily basis. Some of the largest strides have already been made in areas like healthcare where programs utilizing these neural networks can diagnose and even treat illnesses, diseases, and chronic conditions with more accuracy than human doctors. There are also a number of companies making great progress in this area.
A company called ViSEZE has created some applications that use deep learning to create a sophisticated image recognition software. Customers can use images to search for products instead of keywords.
Another company called Atomwise applies deep learning networks to the issue of drug discovery. They use it to explore the possibility of using known and tested drugs for use against new diseases. Also they are looking to create new designer drugs from scratch.
Another fascinating application is being driven by a company called Deep Genomics. They are attempting to use deep learning to predict how natural gene variation changes cellular processes. An example of this being DNA to RNA transcription. Read more about their first product and all of the bold possibilities associated with it here.
These are just a few examples of the many companies who are trying to use this technology to the fullest of its potential.
LOOKING TO THE FUTURE
One of the biggest applications of deep learning which will have a huge impact on our world is better image recognition software leading to self driving cars. Most self driving cars are still having trouble distinguishing between obstacles that require braking versus obstacles that require swerving or obstacles that aren’t really obstacles at all. After that problem is solved it just comes down to some small algorithmic improvements here and there. Google makes some very bold claims about their self driving cars drastically improving within the next few years thanks to effective deep learning models. Tesla also promises huge strides in a short amount of time.
The stage has been set for an AI arms race between tech companies large and small to create everything from self driving cars to chat bots and even self conscious and intelligent programs. Deep learning is now at the forefront of this innovation.