Artificial Neural Networks (ANN) in 4 minutes

The 4 MIN INTUITIONS

About now, 1000s of Sci-fi books and universes of research have been published about how the world would look like in the next centuries. Each, having one element in common. The element of Ai. A more grounded and vanilla version of it has been studied for decades, called Machine Learning. But the thing that provides, exponential potential to this field is Deep Learning. And Artificial Neural Networks?

ANNs are the atom, building blocks of everything we have hypothesized until now about the future. Every algorithm, from terminator to transcendence, fancy Genetic Algorithms are in someway built on ANNs. This post, we’ll get you ANN’s intuition, what it is and how it works (without the generic cliches of it’s history, papers and yada yada yada)

What Are ANNs ?

An attempt at simulating the human brain, ANN provides a structures to replicate the billions of neurons a human brain uses to take any decision.

Think of an if/else condition, returning a boolean output. Now think of the if/else containing an elaborately complex condition. Now think of 1000s of such if/else conditions connected to each other with unique relations. The previous one helped us make a boolean decision by conducting a single operation, this is similar. Just that, this would help us take decisions way more complex and is capable of tweaking it’s conditions enabling it to ‘learn’ from previous attempts.

How does an ANN work?

For readers who have some experience with classical machine learning would know. Every fundamental algorithm (eg. regression), forms some sort of a model with trimmed parameters to perform specific prediction. Now if you take 100s of such models and convert each into a parameter itself to be fed within a larger cumulative model, where each individual model represent a neuron we will simply arrive at Deep Learning or an ANN.

The first stage is to train the ANN. You provide the ANN parameters and then you provide the corresponding answers. During this phase, the network compares its actual output produced with what it was meant to produce — the desired output. The parameters which were previously random are tweaked, one step at a time. This iterates a number of times to get a significant accuracy.

What are the parts of an ANN?

Every ANN contains something called an activation function. They introduce non-linearity and squashes the values into a specific range. Eg, Sigmoid activation function will squash the values within 0 to 1.

A simple ANN will contain 3 fundamental layers.

Input Layer: Contains as much neurons as their are parameters this is where we feed in our data.

Hidden Layer: Contains variable no. of neurons which is often manipulated and tweaked to increase the accuracy of the overall model.

Output Layer: Contains as much neurons as their are final labels (if it’s a classification model) and provides the predicted result.

Applications of ANNs

Every significant Deep learning Algorithm, is somehow built on ANNs. Our ‘brain’ is interestingly divided into 4 core segments. And all the major algorithms, are just attempts at replicating the functions of these differennt parts of our brain.

1. Temporal Lobe -> ANN: This lobe’s responsible for long term memory, since ANNs advocate the concept of ‘weights’ where they will learn something and store them, such that the learned system can be revisited philosophically suggests ANNs mimic the Temporal Lobe of our brain.
2. Occipital Lobe -> CNN: This lobe is responsible for vision and recognition of objects and hence CNN algorithms mimics the function of Occipital Lobe.
3. Frontol Lobe -> RNN: Dealing with short term memory, where it considers the previous couple of things that happened so it can take the next action. RNN algorithms help to mimic this part of human brain. (Frontol Lobe’s also responsible for personality, behavior etc)
4. Parietal Lobe -> ???: This lobe is responsible for sensation, perception and construct a spacial coordination system for space around us and we are yet to create a neural network that can fit in this criteria.