Image Recognition With Deep Learning: Introduction

Deep learning, a branch of artificial intelligence (AI), is ubiquitous these days and powers many aspects of our modern life. Like many others, I am a consumer of this technology, often benefiting from content filtering on social networks, recommendations on e-commerce websites, and enhanced features in smartphones.

So, I started on a journey to learn this technology, beginning with a basic project to build an image recognition model Keras API.

But first, let's learn a little about Deep Learning:

Deep Learning is a subset of AI based on neural network architecture. In fact, AI - and by extension, Deep Learning - are not new concepts. People started to explore neural networks shortly after the invention of computers, which were inspired by the structure and function of human brains.

AI took a long time to become a reality because the neural networks of the early days did not perform as well as the traditional Von Neumann computer architecture, which serves as the basis for modern CPUs.

Von Neumann architecture is based on user-defined limits. Therefore, it has traditionally required an army of engineers to teach a computer what to do by defining strict programming rules. These computers are only as good as the code they run and the engineers who wrote the code.

The Von Neumann architecture works very well for well-defined but limited tasks, but it is not efficient and sometimes fails to perform simple generalized tasks, such as image recognition or sentiment analysis. For example, humans have no problem with the simple generalized task of telling the difference between an image of a cat and a tree.

Turns out that the neural network architecture is much better at performing generalized tasks, such as speech and object recognition. It does this by learning through trial and error, the same way a human child learns the difference between cats and trees.

Neural networks are becoming mainstream now, even though they are as old as the invention of computers, because they need a mountain of data to learn from through 'trial and error.' We now have a vast amount of data available for training Deep Learning models, thanks to the age of the internet. Additionally, computers have become much more powerful and are able to process the vast amount of data needed to train an AI engine.

In the next article, I will use Deep Learning to train a neural network to recognize images of 10 different objects. I will use the CIFAR-10 dataset, which contains 60,000 32x32 color images.