Neural Networks form the foundation for Deep Learning, the technique AlphaGo used with Reinforcement Learning (RL) to beat a Go master. In this article, we’ll explain how the basics of neural networks work.
The focus of this series is to dissect the methods used by DeepMind to develop AlphaGo, the machine learning program that shocked the world by defeating a worldwide Go master. By peeking under the hood of DeepMind’s algorithm, we hope to demystify Machine Learning (ML) and help people understand that ML is merely a computational tool, not a dark art destined to bring about the robot apocalypse. In the earlier articles we discussed why AlphaGo’s victory represents a breakthrough, and we explained the concepts and algorithms behind reinforcement learning—a key component of DeepMind’s program. In this article, we’ll explore artificial neural networks. Neural networks form the foundation of deep learning, the technique that enabled DeepMind’s reinforcement learning algorithm to solve extremely large and complex problems like Go. Deep learning is an advanced form an artificial neural network. So, before we dive into deep learning in the next article, we’ll first explore how a neural network operates.
Neural networks were conceived in the middle of the twentieth century as scientists started understanding how the brain works. What makes the brain so interesting to computer scientists is that in many ways it acts like a mathematical function: input comes in (in the form of electrical and chemical signals), some type of neurological computation happens, and output goes out to other cells. For example, you receive input from your sensory system, perhaps noticing a burning sensation on your hand. Then your brain performs some computation, such as determining that your hand has touched a hot pan. Finally, electrical output is sent to your nervous system causing you to contract your arm muscles and jerk your hand away from the hot surface.