Neural Networks for Dummies: A Comprehensive Guide

Our brain is an incredible pattern-recognizing machine. It processes ‘inputs’ from the outside world, categorizes them (that’s a dog; that’s a slice of pizza; ooh, that’s a bus coming towards me!), and then generates an ‘output’ (petting the dog; the yummy taste of that pizza; getting out of the way of the bus!).

All of this with little conscious effort, almost impulsively. It’s the very same system that senses if someone is mad at us, or involuntarily notices the stop signal as we speed past it. Psychologists call this mode of thinking ‘System 1’, and it includes innate skills — like perception and fear — that we share with other animals. (There’s also a ‘System 2’, to know more about it, check out the extremely informative Thinking, Fast and Slow by Daniel Kahneman).
How is all of this related to Neural Networks, you ask? Wait, we’ll get there in a second.

Look at the image above, just your regular numbers, distorted to help you explain the learning of Neural Networks better. Even looking cursorily, your mind will prompt you with the words “192”.
You surely didn’t go “Ah, that seems like a straight line, I think it’s a 1”. You didn’t compute it – it happened instantly.

Fascinating, right?

There is a very simple reason for this – you’ve come across the digit so many times in your life, that by trial and error, your brain automatically recognizes the digit if you present it with something even remotely close to it.

Let’s cut to the chase.

What exactly is a Neural Network? How does it work?

By definition, a neural network is a system of hardware or softwares, patterned after the working of neurons in the human brain. Basically, it helps computers think and learn like humans. An example will make this clearer:
As a child, if we ever touched a hot coffee mug and it burnt us, we made sure not to touch a hot mug ever again. But did we have any such concept of hurt in our conscience BEFORE we touched it? Not really.
This adjustment of our knowledge and understanding of the world around us is based on recognizing patterns. And, like us, computers, too, learn through the same type of pattern recognition. This learning forms the whole basis of the working of neural networks.

Traditional computer programs work on logic trees – If A happens, then B happens. All the potential outcomes for each of the systems can be preprogrammed. However, this eliminates the scope of flexibility. There’s no learning there.
And that’s where Neural Networks come into the picture! A neural network is built without any specific logic. Essentially, it is a system that is trained to look for and adapt to, patterns within data. It is modeled exactly after how our own brain works. Each neuron (idea) is connected via synapses. Each synapse has a value that represents the probability or likelihood of the connection between two neurons to occur. Take a look at the image below:

What exactly are neurons, you ask?
Simply put, a neuron is just a singular concept. A mug, the colour white, tea -, the burning sensation of touching a hot mug, basically anything. All of these are possible neurons. All of them can be connected, and the strength of their connection is decided by the value of their synapse. Higher the value, better the connection. Let’s see one basic neural network connection to make you understand better:

Each neuron is the node and the lines connecting them are synapses. Synapse value represents the likelihood that one neuron will be found alongside the other. So, it’s pretty clear that the diagram shown in the above image is describing a mug containing coffee, which is white in colour and is extremely hot.

All mugs do not have the properties like the one in question. We can connect many other neurons to the mug. Tea, for example, is likely more common than coffee. The likelihood of two neurons being connected is determined by the strength of the synapse connecting them. Greater the number of hot mugs, the stronger the synapse.
However, in a world where mugs are not used to hold hot beverages, the number of hot mugs would decrease drastically. Incidentally, this decrease would also result in lowering the strength of the synapses connecting mugs to heat.
So,

Becomes

This small and seemingly unimportant description of a mug represents the core construction of neural networks.
We touch a mug kept on a table — we find that it’s hot. It makes us think all mugs are hot. Then, we touch another mug – this time, the one kept on the shelf – it’s not hot at all. We conclude that mugs in the shelf aren’t hot. As we grow, we evolve.

Our brain has been taking in data all this time. This data makes it determine an accurate probability as to whether or not the mug we’re about to touch will be hot. Neural Networks learn in the exact same way.
Now, let’s talk a bit aboutthe first and the most basic model of a neural network: The Perceptron!

What is a Perceptron?

A perceptron is the most basic model of a neural network. It takes multiple binary inputs: x1, x2, …, and produces a single binary output.

Let’s understand the above neural network better with the help of an analogy.
Say you walk to work. Your decision of going to work is based on two factors majorly: the weather, and whether it is a weekday or not. The weather factor is still manageable, but working on weekends is a big no! Since we have to work with binary inputs, let’s propose the conditions as yes or no questions. Is the weather fine? 1 for yes, 0 for no. Is it a weekday? 1 yes, 0 no.

Remember, we cannot explicitly tell the neural network these conditions; it’ll have to learn them for itself. How will it decide the priority of these factors while making a decision? By using something known as “weights”. Weights are just a numerical representation of the preferences. A higher weight will make the neural network consider that input at a higher priority than the others. This is represented by the w1, w2…in the flowchart above.

“Okay, this is all pretty fascinating, but where do Neural Networks find work in a practical scenario?”

Real-life applications of Neural Networks

If you haven’t yet figured it out, then here it is, a neural network can do pretty much everything as long as you’re able to get enough data and an efficient machine to get the right parameters. Anything that even remotely requires machine learning turns to neural networks for help. Deep learning is another domain that makes extensive use of neural networks. It is one of the many machine learning algorithms that enables a computer to perform a plethora of tasks such as classification, clustering, or prediction.

• With the help of neural networks, we can find the solution of such problems for which a traditional-algorithmic method is expensive or does not exist.
• Neural networks can learn by example, hence, we do not need to program it to a  large extent.
• Neural networks are accurate and significantly faster than conventional speeds.

Because of the reasons mentioned above and more, Deep Learning, by making use of Neural Networks, finds extensive use in the following areas:

• Speech recognition: Take the example of Amazon Echo Dot – magic speakers that allow you to order food, get news and weather updates, or simply buy something online just by talking it out.
• Handwriting recognition: Neural networks can be trained to understand the patterns in somebody’s handwriting. Have a look at Google’s Handwriting Input application – which makes use of handwriting recognition to seamlessly convert your scribbles into meaningful texts.
• Face recognition: From improving the security on your phone (Face ID) to the super-cool Snapchat filters – face recognition is everywhere. If you’ve ever uploaded a photo on Facebook and were asked to tag the people in your photo, you know what face recognition is!
• Providing artificial intelligence in games: If you’ve ever played chess against a computer, you already know how artificial intelligence powers games and game development. It’s to the extent that players use AI to improve upon their tactics and try their strategies first-hand.

In Conclusion…
Neural networks form the backbone of almost every big technology or invention you see today. It’s only fair to say that imagining deep/machine learning without neural networks is next to impossible. Depending on the way you implement a network and the kind of learning you put to use, you can achieve a lot out of a neural network, as compared to a traditional computer system.

Learn ML courses from the World’s top Universities. Earn Masters, Executive PGP, or Advanced Certificate Programs to fast-track your career.