Recent developments in artificial intelligence (AI) are now applying advanced computational methods, called deep learning, to new applications. These applications range from predicting what kind of movie you would enjoy or how well someone could do their job based on their resume to creating automated programs that can mimic human thinking.
Deep neural networks have been used as the basis for many different types of AI since they were first introduced in the 1980s. They work by using multiple layers of algorithms to process data sequentially. As these systems get smarter, though, researchers must be careful not to overtrain the system- it will become too dependent on the input data and lose its ability to generalize outside of those patterns!
There are some who use the term “machine learning” broadly to include all areas where computers apply learned information to solve problems, but this is definitely not the accepted usage of the term. In fact, there is an effort underway to create an alternative nomenclature for technologies like deep learning.
This article will focus only on discussions surrounding the relation between deep learning and machine learning, so feel free to skip ahead if you want more info about the other terms. For additional reading, check out our article: What is Neural Networking? And our article: The Differences Between Artificial Intelligence and Machine Learning.
Definition of deep learning
Definitions of deep learning vary slightly, but typically refer to algorithms that use neural networks to perform specific tasks. Neural networks are complex biological structures in our brain that help us process information through repetition.
Deep learning is just such an algorithm that uses neural networks to automate some processes or functions. For example, it can be used to recognize objects, classify images, predict outcomes, and so on.
There’s no singular definition, however, most agree that deep learning involves complicated math equations trained with lots of data! That last part is important–it takes a lot of data to make the machine learn how to apply the theory correctly.
And we need to emphasize that this mathematical equation isn’t simply repeated over and over like other algorithmic concepts (like those mentioned above). No, these equations evolve as they are fed more and more data, creating increasingly sophisticated patterns. In fact, one could say that the mathematics become indistinguishable from the pattern itself at that point.
History of machine learning
While not widely known, there is an older theory behind what makes AI work so well these days- deep neural networks!
Deep neural networks are not new, but they have become more popular in recent years as their use has exploded across industries.
By incorporating features into very complex mathematical functions, you can create algorithms that learn patterns and correlations from data. This technology was first theorized by Warren McCulloch and Walter Hecht back in 1959!
McCulloch and Hecht called this concept “neural nets” or “feedforward nets,” which refer to how the network learns.
These concepts influenced later theories such as recurrent neural networks (RNNs) and long short term memory (LSTM).
However, it took computer scientists decades to develop machines with enough computational power to run large neural net models. That all changed around 2010 when GPUs made developing RNNs and LSTMs fast and easy.
History of deep learning
Recent developments in artificial intelligence (AI) have been referred to as “deep learning” or, more commonly, just “neural networks.” This is due to the fact that these systems use computational architectures modeled after how neurons work in our brains!
Neurons are thought to play an important role in shaping our perception of the world around us. For example, when you see someone walking away, your brain automatically form a hypothesis about their destination based on their body language and other information it has gathered.
In neural network parlance, this process is called pattern recognition or feature extraction.
By using such features to compare with similar patterns from past experiences, the net can determine what the person was going to do next.
With enough examples, the system learns what normal human behavior looks like for each element of the trip, and uses that knowledge to make predictions.
A key part of this analogy is that there are no pre-defined rules — every neuron connects to every other one, creating a complex interplay between input and output.
This architecture allows neural nets to find increasingly sophisticated ways to match up inputs and outputs. Some say it resembles the way humans learn and apply knowledge.
Experts agree that AI will soon achieve superintelligence, which is defined by Elon Musk as having “achieved levels of self awareness and cognitive function far beyond those we currently give it credit for.
Differences between machine learning and deep learning
While both of these terms are typically grouped together, there is some slight difference in what they refer to. Technically speaking, a term like “machine learning” can be interpreted as referring to either or neither. Some people use only the term “deep learning” for things that contain an artificial neural network!
With that out of the way, let’s dive into the differences.
Deep learning is usually characterized by two things: advanced layers and lots of data. As you’ll see below, this distinction makes sense because it tells us something about how each technique works.
Applications of machine learning
Recent developments in artificial intelligence have come under the category of “machine learning.” This is a field that uses computers to learn tasks or patterns, then use those learned lessons to perform new tasks.
Machine learning has applications in almost every area of technology. It can be applied to recognize handwriting, decipher images and videos, understand spoken language, predict future behavior, and more.
Deep neural networks are one type of machine learning algorithm used in computer science and engineering. They became popular around five years ago when engineers started applying them to solve complex problems in perception (seeing) and action (doing).
Since then, they have become integral parts of other types of algorithms, making it possible to apply deep learning to ever-more difficult challenges. That’s why most people associate the term “deep learning” with AI now.
Computer scientists have also experimented with adding additional layers to the network architecture to see how well different configurations work. A nested structure called “deep belief network” was found to outperform earlier architectures like convolutional nets and recurrent neural networks.
These theories inspired the creation of newer generations of neural networks.
Applications of deep learning
Recent developments in artificial intelligence (AI) have been focused largely on two things: applications that use computer programs informed by mathematical theories to perform specific tasks, and systems that use neural networks or other algorithms inspired from how humans process information to learn new skills or redefine what those skills are.
The first type is referred to as “application-specific” AI or intelligent software. Systems like Amazon’s Alexa can respond to questions using advanced natural language processing, for example. Software capable of performing image recognition such as Google Cloud Vision and Microsoft Azure Computer Vision are growing in popularity.
And while these types of systems still mostly rely on programming, they go beyond it in terms of what they can be programmed to do. These so-called generalist AIs can be taught about objects, pictures, speech, and more, which gives them broader functionality than simpler, programmatic AIs.
Generalist AIs, however, don’t work without large amounts of data to train on. That’s where the second branch of AI comes into play: machine learning.%CLB
Unlike generalized AIs that require lots of training material, ML techniques aim to teach themselves how to operate through repeated exposure to datasets that feature examples of the concept being learned. For instance, an algorithm looking to identify cats may test many different cat photos against a definition of “cat.” If there’s enough instances of both, then the system would likely consider that photo to be a feline.
What is the future of machine learning?
Recent developments in artificial intelligence (AI) have been referred to as “machine learning” or, more commonly these days, just “deep learning.” Both terms refer to algorithms that enable computers to learn tasks for themselves, something we humans are naturally capable of!
Deep neural networks are very complex systems that process large amounts of data using concepts called layers to make predictions. For example, when you look at a cat picture, it is not enough to know that there is a shape with fur and whiskers — what kind of furry white shape has whiskers?
A deep convolutional network will compare all possible shapes with respect to those features and choose which one resembles the most a real cat. Algorithms like this allow machines to automatically find patterns in vast expanses of information.
It is no surprise then that people are investing heavily into AI-powered technology. Companies such as Google, Facebook, Amazon, Microsoft, and Apple use it in various ways to improve their products and services.
What is the future of deep learning?
Recent developments in neural network architectures have led to what many refer to as “deep” or more advanced versions of machine learning. These newer networks typically use multiple layers that are connected together, enabling the system to learn increasingly complex patterns.
Deep learning has seen significant success in applications such as computer vision (where it was used to recognize objects, letters, and numbers) and natural language processing (for example, reading texts and understanding their content). It also plays an important role in speech recognition and self-driving cars!
However, one field where people are not quite sure how well deep learning will work is medical image diagnosis. That’s why we’re exploring different types of deep learning for this application here at Stanford.