The human brain is quite magnificent.
As children, we hardly know a thing until we explore, make mistakes, and learn from them. As teenagers, we continue to make mistakes, but previous experiences help shape future choices. As adults, we have so much experience, that we become sharp and refined. Tasks that were once difficult for us have now become routine.
This is all done with the help of our brain taking in context, building vast networks, and tapping these networks to take calculated next steps. In a way, this gives us some perspective as to how deep learning works.
What is deep learning?
Learning by example may come naturally to humans, but not for machines. It requires deep learning to complete tasks that are second nature to us.
What is meant by deep learning?
Deep learning is a subset of machine learning. It uses artificial neural networks with many hidden layers to extract features from raw data. The more input data, the more the model learns.
If this sounds awfully similar to standard machine learning, it’s because deep learning is machine learning – just with different applications. Here’s how it differs.
Standard machine learning, like supervised and unsupervised learning, works in a few ways.
- Uses training data from sample datasets. For example, pictures of cats.
- Labels the data to provide context. Cat = 1, not cat = 0
- Requires an output value, like being able to discern cats from non-cats.
- Uses a machine learning algorithm to map input and output values.
- Validates the model using more labeled data. This is for accuracy testing.
Supervised learning is commonly used to help email platforms discern traits of spam mail from regular mail.
- Uses no training data.
- Uses no labels, nor context.
- Instead, uses a machine learning algorithm to find naturally occurring patterns in large datasets.
Unsupervised learning is commonly used to identify suspicious behavior when it comes to detecting and preventing financial fraud.
Ok, so how is deep learning different?
The key difference separating deep learning from standard machine learning methods is the use of artificial neural network software. These are often referred to as deep neural networks and are in some ways similar to how the human brain processes information.
Let’s break down deep learning and the inner-workings of these networks.
Deep learning first requires massive amounts of labeled data – far greater than what you’d see in standard machine learning methods. This data could be millions of images and lines of text or thousands of hours of video footage.
Because so much data is being processed, this requires some hefty computing power. Hence, why deep learning hasn’t really made many strides until recently.
Then comes the neural network.
Artificial neural networks, they’re like brains, but not really
Because deep learning is still a subjective term, I asked Brian McGuckin, one of our data scientists with deep learning knowledge, to define neural networks.
“Historically, neural networks have been described as attempting to simulate human brain neurons, though this became somewhat of a crutch and incorrectly implies that we know and can model how humans learn (we don’t, so we can’t). Neural networks are basically an aggregation binary classification algorithms known as nodes (or neurons) and organized into layers.
At a minimum, a neural net has two layers: a ‘hidden’ layer with many nodes, and a single node ‘output’ layer which is what translates complex calculations into our final predictions. Layers are referred to as hidden since we don’t see exactly what, how, and where certain calculations are performed (though research on this topic exists and is picking up momentum). This two-layer network is referred to as ‘shallow’. A deep neural net is then any neural net with more than one hidden layer.”
In other words, multiple layers of neural nets is what defines a deep neural network and differentiate deep learning from standard machine learning.
Some neural networks are wildly intricate, having as many as 150 hidden layers. Each layer plays a key role in breaking down features of raw data. This is formally called feature extraction.
Deep neural networks with few layers are actually used across banks and post offices today to recognize handwriting styles. This comes in handy when cashing checks with your mobile phone.
Because handwriting styles are unique to each of us, it requires a lot of training data for machines to recognize numbers and letters. It’s safe to say the technology has reached a point where it is reliably accurate. This, however, couldn’t be possible without deep learning.
More complex neural networks with many layers are being used right now to develop driverless cars.
Source: Fred Lambert, Electrek
Extracting features from the road, recognizing crosswalks and traffic signs, and understanding movement patterns of other vehicles are just a few of the many types of raw data being broken down piece-by-piece in deep neural networks.
Fun Fact: Have you ever completed a CAPTCHA while using Google? Having to select squares of signs and roads? Well, Google actually uses this insight to train its deep neural networks for driverless cars. Thanks for helping out!
The advancement of learning
Machine learning has set the stage for deep learning, and neural networks will help build more human-like artificial intelligence of the future. With more time and resources, the applications of deep learning are nearly endless.
A disruptive type of deep learning, called reinforcement learning, is making headway in gaming and robotics, but will soon touch other industries. See how positive and negative reinforcement is being paired with complex algorithms.