Layman Explanation of Artificial Neural Networks
Updated: Dec 18, 2021
For a more technical explanation on neural networks, refer to one of my older posts.
With the explosion of AI across all industries, Deep Learning has been responsible for many of the major breakthroughs. But for people in non-tech, what exactly is Deep Learning?
Deep Learning is a series of machine learning models based on artificial neural networks.
Okay, we just described what is Deep Learning. However, the definition doesn’t define what exactly is an artificial neural network (ANN).
Before going into what exactly is an ANN, it would be best to first go into how the human brain functions.
Our brain is a very powerful machine. It enables for movement and processes information from one’s surroundings. In addition, our brain allows us to learn new things and habits .
All these things are possible due to the brain consisting of billions of neurons sending information from one another. Many neurons are required in order to perform tasks. Below represent networks formed by several neurons.
Now, how exactly is information transmitted from one neuron to another?
Take our neuron below.
The branches on the left represent dendrites. The dendrites are responsible for taking in chemical reactions (information) from other neurons and transmitting them to the other end. At this point, information is sent to other neurons.
Of course, not all neurons are fired up whenever you want to do a particular task. Instead, neurons will only fire up if a certain threshold is reached. This is known as action potential.
Representing the Artificial Neuron
Now how does the above analogy tie into artificial neural networks? Good question.
Just like a biological neural network, an artificial neural network consist of a series of artificial neurons to make decisions.
As a note, an artificial neuron is better known as a perceptron in the AI field. So how exactly does this perceptron (artificial neuron) work? Take the image below.
A basic structure of an artificial neuron.
Unlike an actual neuron, a perceptron uses math to calculate an output.
In the image above, a series of inputs are provided to the perceptron. The perceptron then multiples the inputs with its weights to get an intermediate output. With the intermediate output, we send it through a function that’ll return a number ranging from 0 to 1 (or some other step function). This function is called the activation function and serves the biological neuron’s equivalent of action potential.
Translating the image above, we get the following equation for a perceptron:
For those familiar with statistics, this equation looks similar to another machine learning algorithm called logistic regression. In fact, the equation to a perceptron is very similar to that of logistic regression.
From the previous section, perceptrons are just logistic regression. While they can somewhat handle nonlinear datasets, they cannot handle complex datasets, like images. So now what?
This problem was encountered shortly after researchers concluded that perceptrons by themselves were too weak for complex datasets. It was discovered that using multiple perceptrons together can model more complex datasets. Then taking a series of perceptrons, we can form a layer. A layer then serves to compute a series of inputs and outputs. You can even stack layers together to create even more complex models. As a result, your model looks like a neural network.
A neural network with 3 layers
Like all things in life, there are several types of artificial neurons that can be used to represent very complex models. Going through each type is beyond the scope of this post.
Where to Go From Here
ANNs serve as the backbone for many deep learning concepts and applications across many fields. Now where to go from here?
There are a few way to go about this:
If you are not interested in learning more about neural networks (and deep learning by extension), hopefully this post gave you an idea on what are neural networks.
If you are interested in neural networks but only in applying it to your industry, read up on the latest news on how they’re being implemented in the industry. If possible, bring in experts in deep learning to help explain and implement your idea.
If you want to implement neural networks personally but not interested in the theoretical aspect, learn more about the types of neural networks as well as frameworks to implement them. Tensorflow 2.0 was recently released and allows users to easily create complex neural networks.
If you’re interested in advancing the deep learning and AI field in general, you’d be best off going to grad school or get a PhD in AI.