The Science, Architecture and Advantages of Neural Networks Revolutionizing Artificial Intelligence

Artificial Intelligence (AI) is a broad field, which aims at making machines intelligent. It encompasses a set of tools through which it enables a machine to mimic human intelligence.

One of the tools AI has is Machine Learning, which enables the machines to learn without being told explicitly what to do.

Machine learning again has a portfolio of tools, one of them being neural networks. Neural Networks somewhat replicate or mimic the activity of a human brain. Deep learning is the use of more sophisticated neural networks, with more non-linear layers, convolutional layers, etc.

Today, neural networks and neurocomputing have revolutionized artificial intelligence (AI) and made exponential advances possible in deep learning. Their utility also extends into natural language processing, speech recognition and computer vision, the very purpose of the original Perceptron, i.e., a computer model or computerized machine devised to represent or simulate the ability of the brain to recognize and discriminate.

Let’s take a deep insight into the science, architecture, and advantages of neural networks, all of which start with how they emulate the human brain.

Neural networks are computer systems modeled on the human brain and nervous system. They are also known as artificial neural networks or neural nets. Their solid foundation in computer science draws roots from neurology and the flourishing field of neuroplasticity in particular. Popularized by books such as The Brain That Changes Itself by Dr. Norman Doidge, neuroplasticity revolves around the theory that brains can change as a person’s circumstances do, say for example, from brain haemorrhage trauma, shocking incident or redundant exposure to a stimulus.

Now researchers in neuroplasticity contend that the concept of a hard-wired brain, one that is constant and unchanging in its makeup, is inaccurate. Neurons that fire together wire together, is a phrase that was coined to explain how we learn and retain knowledge over time. In a manner that is similar to the neuroplastic human brain, neural networks question the idea of hard-wired computing, since they can learn without any prior knowledge of how to perform a task.

Imagine an operations executive, who learns to perform an iterative task through exposure to patterns, practice and an increasing sequence of efficient procedures. All of this over time becomes imprinted on her firing neurons. Now, think of a computer that can recognize patterns through what are called training examples, for example, millions of images fed into a neural network that learns to distinguish among a pen, a car, a horse and a mug. The network then sharpens its knowledge while it performs its assigned tasks in real time, a procedure known as deep learning.

Neural networks function as software simulations of the brain. Therefore, an artificial neural network (ANN) isn’t a structure, rather a computer program or algorithm that organizes the billions of transistors in a computer to operate as though they were interconnected neurons in a human brain.

In the simplest terminology, artificial neurons, also referred to as nodes or units, are mathematical functions carried out by transistors and take their directions from the program. These neurons are organized into three types of layers, namely input, hidden and output. The key to understanding neural networks is by getting a hang of the flow of how they process information. As an example, let’s use a network programmed to recognize a handwritten digit from zero to nine.

Input layers receive the information from an external source, which in this case, is the visual image. Consider plugging a camera into a computer that has been programmed as a neural network. The camera’s image would first hit the input layer. No actual computation is taking place here. The input units are simply receiving information and doing nothing more. So if the image of a number is broken down into 1024 pixels (32 by 32), the input layer is simply communicating with the network, thereby informing which pixels are lit up, and which ones aren’t.  

This piece of information is then passed onto Hidden layers, so called because they aren’t connected to the external world, the way inputs and outputs are. Here, the input information gets defined more precisely with every hidden layer it passes through. Just like neurons in the human brain, neurons in the hidden layer fire, and based on how they fire, neurons in the next layer fire as a consequence.

In our case under consideration, first hidden layer might determine whether the lit pixels are organized into edges; second layer, whether the edges are organized into patterns; third layer, whether those patterns are straight or looped; and so on.

Finally, the Output layer is where the network gives us the final result, a number from zero to nine.

This entire process which can be visualized as a left-to-right workflow, is known as a feed-forward network. But what if you want your neural network to learn, that is, to keep refining its outputs until it gets faster, efficient and more accurate at what it does? That necessitates taking the network’s output, comparing it with a standard result, and feeding it back into the network from scratch. This entire feedback-loops process is known as backpropagation.

The revolutionary discovery of backpropagation came in 1986, when Geoffrey E. Hinton, a professor at Carnegie Mellon University, became one of the first researchers to define what he called learning procedures, appropriately explained in his seminal paper as methodologies computers could learn by performing a task over and over, each time with the computer’s neural network then adjusted in the direction that diminishes the error.

Therefore, a computer receives an input, processes it through hidden layers, outputs it, and through a backpropagation algorithm re-inputs it to refine its performance and update its knowledge.

The power of neural networks comprise of speed, scale, and accuracy. It signifies the key advantage that neural networks provide. Consider banking services, where millions of credit card transactions per minute pass through a computer system. A neural network can not only keep up, but it can also flag potentially fraudulent transactions, based on a number of input variables, and with fewer false positives. Refer one of my previous posts here. The variables might range from the number of transactions over a short period to whether the card is being used in an unusual location or sudden spike in usage frequency. The case would then be handed over to a human bank official for further investigation, with the card temporarily frozen to protect the consumer, who is then notified.

Neural networks can also learn and calibrate fast enough to handle a host of ever-changing situations. A few classic examples of this include the airplane autopilot that applies course correction on need basis and the current development of self-driving autonomous cars.

The current scenario is that neural networks cannot rival the number of connections in the human brain, which some estimate as at least a hundred trillion, could be more. However, the way artificial intelligence is being embraced by innovative technology leaders like Google, it tells us something about where neural network technology is headed. We should be looking forward to exponential growth in processing power, speed and learning ability over the next five to ten years.