We are drowning in information but starved for knowledge.

- John Naisbitt

Follow us





The Daily Feed


Neural Networks

Neural Networks Breakdown Part I

163

Neural networks are the heart of most of the existing AI systems we have at our disposal. Neural networks allow us to create facial recognition, image classifiers, speech recognition, language translation, etc. Let us see how it does it. <h2><b>Visualization</b></h2> A neural network is just a bunch of nodes stacked in layers. That is too simple, right? Yes, but this is what it actually is. Knowing this, let us try to visualize a neural network. <img src="/static/nn.png"></img> The nodes are actually stacked into layers. The first layer is called the input layer and the final layer, with just one node is called the output layer. The "y" is called the output and the x<sub>1</sub>, x<sub>2</sub>, x<sub>3</sub> and x<sub>4</sub> constitute the input signals or just inputs, for the neural network. Each layer is connected to the next layer through connections which have specific weights. These weights are used to generate the outputs for each node. The output is calculated through the summation of the multiplication of weight and the corresponding input. This value is then sent through an activation function, to give us the output. Through this, we have constructed a quick visual of how a neural network looks like. <h2><b>Specifics</b></h2> The activation function is what gives the neural network its significance. This function takes the summed up value and transforms into a value between a smaller range. Generally, the activation functions help us get the value in between the range of 0 and 1. This is analogous to the neuron which has a potential of either firing or not. The activation function too, gives a value near to 1, indicating "firing", or 0, indicating "not firing". Let us say we want to identify whether an animal is a dog or not. Well, for a human, it is fairly easy to decide, as he or she knows what a dog looks like. We know the specific features which a dog has and others don't, which classify it as a dog. How do we know that though? For many years, we have been fed the "training data", which helps us know which animal is which. When we come across something we've never seen before, we can't identify what animal that is. For training a neural network, we have to do the same. We have to feed the neural network data, for features which correspond to the dog. We also send it with labels, giving a 0, for the values which don't correspond to a dog and 1 for those which correspond to a dog. When the neural network is trained, it can tell which values identify a dog and which don't. Just like a human can do, when it knows which animal is which. It will return a value near to 0 or near to 1, for images which are not a dog and for those which are, respectively. Now it may be clear, why the activation function is so crucial, and how a neural network can act as a "dog" classifier. We will continue the discussion <a href="/Neural-Networks-Breakdown-Part-II">here</a>.

- Shubham Anuraj, 12:28PM, 16 July, 2018

AI