Deep learning sounds so cool, doesn’t it? There is machine learning, which is already super cool and now make it deep. Even better.
But what makes a learning algorithm deep?
Even though the term is super high fashion to use and is being thrown around like a magical solution to all our problems, it actually just means that we stack things on top of each other.
Deep learning models on the base level, consist of neurons. A neuron is very similar to a logistic regression model. By stacking neurons on top of each other to make layers and then stacking those layers one after another you make a deep neural network.
Of course, I’m being overly simplistic here. But on a high level that’s what a deep learning algorithm is.
(Quick note here: deep learning is a type of machine learning. So actually the term machine learning contains deep learning. But in the context of this article, when I say machine learning, I mean algorithms that are not part of deep learning.)
Machine learning algorithms have been around and part of life since the early 90s. Whereas deep learning algorithms have not been used until very recently, that is 10-12 years ago. Back then, our computers were not strong enough to run deep neural networks. In the meantime, we came up with better processors and a bit more efficient calculations to make them feasible to train in a short amount of time.
Okay so now we know the infrastructural difference but how are they different from ML algorithms when it comes to learning and results?
One gigantic advantage DL models have is that they do not need features to be prepared for them.
If we want to make an image annotator using random forests let’s say, we would have to first extract some features from the images. This could be the amount of pixels that are green, the amount that are blue, how many people there are in the images, certain corners and their angles etc. But if we want to do the same with a deep learning model such as a CNN, all we need to do is to provide the image as it is to the model. It learns where to look itself. So there is no feature engineering needed.
It sounds like a dream, I know…
But it comes with a price. You need much more data points to make the model accurate. And generally stronger computing power to be able to train because of the complexity of the calculations being done to train the model.
But this characteristic makes the DL model be able to solve problems we wouldn’t know how to frame.
So now you know what to say if someone asks you how machine learning and deep learning are different. If you want to learn how deep learning works and how to implement it yourself, make sure to sign up for the updates of my upcoming course. You can read about the course and check out the preliminary table of contents here.