Sorry, you need to enable JavaScript to visit this website.

Neural Networks: Inception and Insights

Neural Networks: Inception and Insights
April 05, 2019

They say space is the final frontier, which may be true to some extent, but the gateway to the final frontier lies far closer to one’s own home than what we are led to believe. The human brain, an object of fascination for the scientists of the world, is that elusive boundary.

Not just scientists and neurologists, but every human being has, at some point, thought of the infinite mysteries that reside inside their own heads. They say space is the final frontier, which may be true to some extent, but the gateway to the final frontier lies far closer to one’s own home than what we are led to believe. It is interesting that both question and answer originate from the same place that paved the way to the inception of one of the most groundbreaking works that have since changed how we perceive computers. Or rather, how computers perceive us.

The concept of neural networks is derived from a basic mapping of how human and animal brains perceive data. Neural networks may be just vaguely adapted off from brains, but they have created a whole avenue of algorithms that can increase computational power, prowess, and efficiency. Each node in the algorithm represents a neuron and the flow of information between the nodes mapped after the flow of synapses inside the living brain.

There is no single ‘neural networks algorithm,’ but it is more of a collection of frameworks that has been built to congregate machine learning algorithms and, thus, process heavy loads of data.

There are several types of artificial neural networks (ANN). Let us review some of them. Artificial neural network or ANN can be used for just about anything that requires a computer to parse through large sets of data, process them as required, and then act on the results of the information outcome.

In Bottleneck Neural Networks, each subsequent layer reduces the number of nodes, thus creating a bottleneck effect and reducing the amount of data to be processed. They are used for work, such as image compression or data compression.

Neural networks have a wide range of application in computer vision, speech processing, diagnosis, and other decision-making. Technology companies like Google, Adobe, Facebook, and Instagram use this framework widely in their products. A futuristic vision is that a computer should be able to understand and interpret data based on images and videos fed through neural networks.

Neural networks have a wide range of application in computer vision, speech processing, diagnosis, and other decision- making activities.

Artificial neural network or ANN can also be used in the gaming field. The most iconic example of ANN being used for gaming is when a developer creates an AI using neural networks to beat Google Chrome’s offline Dino run game. The work done is amazing, not because of its difficulty, but more so because it shows that ANN can be used for any field, making everything it is applied to work in a simpler way.

The most simple and straightforward type is a Feedforward Neural Network which consists of nodes that feed information to the successive node and so on until the final node is reached. These are used in X-ray image fusion, a process of overlaying two or more images based on the edges.

Recurrent Neural Networks use memory concepts that enable neurons to act like memory cells while performing operations. A common instance where a recurrent neural network is deployed is in text-to-speech conversion.

Convolutional Neural Networks are much like feedforward systems in the sense that they use calculable and updated weights and learning rates and, in addition, use batch processing to open new doors in the field of images and signal processing. This can be used in the field of agriculture to predict future growth and yield of a particular plot of land while combining weather data. Technology companies can provide the necessary infrastructure to help realise that convolutional neural networks find widespread adoption.

One particular type of neural network which amazed me personally is Generative Adversarial Network (GAN). It is a modular neural network which consists of a ‘generative’ and a ‘discriminative’ neural network. Generative adversarial networks are a class of unsupervised machine learning algorithms wherein the two aforementioned modules are pitted against each other in a winner-based result selection method.

A GAN system is basically a robot artist that can be trained to create images, videos, etc. that are astoundingly similar to things one views in real life. What one must understand is that we do not feed an image and ask a GAN system to choose it or select it and give it back to us. A GAN system is capable of generating its own image to satisfy our level of clarity in what the image must be. They can be taught to create which seems to be the most astounding achievement that machine learning and artificial neural networks have produced.

A portrait generated by a GAN system was sold for $432,000. This heralds the inception of AI creating art that is appreciable even by human art connoisseur standards.

In a GAN system, the discriminative network is one that classifies data based on a few predetermined discriminative factors (e.g.: email is spam vs email is not spam, based on a few choice keywords). It maps these features to labels.

The generative part of the network makes features given a set of labels. It tries to create and mimic the particular features and tries to fool the discriminative network.

This is how GANs work: The generative network creates the requested set of features and runs it through the discriminative network. The discriminative network does its best to prevent fakes or unrequited data generated from being passed. So, the generative network becomes smarter each time until the point where it can successfully pass the tests of the discriminative network. The true genius of this idea lies in the successful pitting of network against network and letting a computer learn from itself.

There is so much more to this world of artificial intelligence, ANN, and machine learning. The more we realize the commonalities between each of these terms and their conjoined profoundness, the sooner we can exploit their full potential and ultimately we might even end up finding the fifth force of nature.