Deep learning models are the typical models of this big data era. They are data hungry; nowadays, there is tons of data to be processed. Also, they are computational requesting compared to the traditional linear regression models; older computers were not able to train useful networks. They only turned out to be feasible after the development of the first backpropagation algorithms, along with some computational advancements.
One of the first NN breakthroughs that earned great respect in the community was the outstanding performance achieved by a Deep Convolutional Neural Network during the LSVRC-2010 contest. Such a contest aimed to classify 1.2 million high-resolution images into 1,000 different classes.
Back in those days, such a model took around a week to be trained. With the advent of cloud computing, along with hardware, algorithms, and software improvements, in the early days of 2018, the same model could be trained through an inexpensive cloud service within a couple of hours. Industries and universities started to trust these models a lot more. A hype arose around AI. Deep learning became a buzzword.
PwC consulting estimates show that up to 2030, AI will contribute up to 13 trillion dollars to the world's economy by augmenting production and consumption. As a cutting-edge technique, NNs and their deep learning models are among the core gears that enable this whole movement.
NNs are likely to greatly impact the way that we produce, build, and consume. Social relations might also be affected by NNs. Universities and non-profit organizations are using it to protect the environment and demand better policies. Indeed, to volunteer for a non-profit organization is a great way to give your data science career a boost. Even better would be to proactively found an organization of your own.
Here are some general characteristics of NNs:
- Unsupervised or supervised learning techniques are available
- Regression and classification tasks can be performed with them
- By employing enough hidden layers and nodes, it's possible for the NNs to approximate any sort of function (linear or non-linear); neural nets are known as universal approximators
- General tasks that they are known to perform well include fraud detection, natural language processing, and face and voice recognition
To mention practical enterprises, TensorFlow, which is Google's open source project for deep learning and one of the engines featured by Keras, has a lot of successful history built on top of it. It has been used to diagnosis diabetical retinopathy, monitor maritime life, and is behind the smart reply laid out by Gmail.
NNs are truly unique and marvelous. Nowadays, these models stand somewhere between science and magic. Sure, there are downsides (more about them later), but given their capabilities and brilliance, it does not sound like a surprise to me that neural nets and deep learning already participate so much in our daily lives.
This brief section's intention was to reinforce the influence and capability that NNs have. In the next one, you will see why every deep learning model is an NN, but not every NN is a deep learning model. Understanding NNs in more detail might be helpful to learn the dos and don'ts of these models; that is what the next section is aimed at.