Making Clear the Difference Between Machine Learning (ML) and Deep Learning (DL)
Artificial intelligence is any technique that enables machines — computers, in particular — to mimic human behaviour and perform similar tasks. Most software could fall under this broad definition. Ultimately, the software intermediates as an agent between us and our objectives, namely to buy online, register a warehouse movement, or study. If such software does not exist, another human agent should step forward to replace it. Then we should instead meet a commercial agent, a logistics manager, or a teacher of the desired subject.
With this definition in mind, let’s first clarify what machine learning is.
What is machine learning?
The first definitions of machine learning included a differentiating element with respect to our first definition of artificial intelligence: the improvement of the system through experience. That kind of software should be able to improve your performance on a task through experience performing that task in an iterative process.
The problems that artificial intelligence systems solve are limited to two main types: classification problems, in which we try to predict discrete responses, for example, a grouping in families — clustering — of data in disjoint sets; or regression problems, where we predict continuous responses, for example calculating the optimal value for a certain action, which can be included in a continuous range of values.
The first definitions of machine learning included a differentiating element with respect to our first definition of artificial intelligence: the improvement of the system through experience. That kind of software should be able to improve your performance on a task through experience performing that task in an iterative process.
The problems that artificial intelligence systems solve are limited to two main types: classification problems, in which we try to predict discrete responses, for example, a grouping in families — clustering — of data in disjoint sets; or regression problems, where we predict continuous responses, for example calculating the optimal value for a certain action, which can be included in a continuous range of values.
The ways in which we humans teach machines to “learn through experience” can be summarized in three: supervised learning; unsupervised learning; and reinforcement learning.
In supervised learning, we humans teach machines what should they know. For example, if we are tackling an animal classification problem, we will label images as lions or elephants. In unsupervised learning, no human intervention occurs in the data processing. One notable example was Microsoft’s famous twitter bot Tay, which had to be unplugged after just 24 hours of operation for his misogynistic and ideological comments. We typically use it to discover underlying structures that we do not know about in our data sets. That is, in that case, we use machines to learn from their conclusions. Reinforcement learning is something similar to what we do when we train our pets: when they perform as expected, the algorithms receive a reward.
Note that the problems mentioned above can be tackled with all three techniques, in particular, the two antagonistic ones: supervised or unsupervised — since reinforcement in the background is a form of supervision. When we attack classification problems with supervised learning, we usually want the machine to learn to distinguish things that we already know, with the aim that it later does it for us, for example, artificial vision systems that recognize objects in a warehouse. It is possible to see real cases of very precise image recognition — textures, shapes, objects — on the IBM Watson website (https://visual-recognition-code-pattern.ng.bluemix.net/) or using apps like Vivino that allows us to know about a bottle of wine simply by taking a picture of its label. In the case of unsupervised learning, what we want is to discover structures that we may be missing. When we talk about continuous problems, with supervised learning we will be able to teach the machine a linear adjustment or formula so that from certain variables it is able to predict others — hence, for example, its use to find price optimals. Unsupervised learning in the continuous domain is often used for the so-called dimensionality reduction — that is, to teach us to display data in a simplified way. For example, cleaning variables from a model that has no correlation with what we are looking for.
Other notable day-to-day examples where machine learning techniques are applied are recommender algorithms — such as those of Netflix or Amazon — or real-time prediction systems of the fastest route based on existing traffic, as occurs in the Waze and Google Maps apps.
So what is the difference with Deep Leaning?
Well, so far, everything we have talked about has been related to machine learning. What then is deep learning or deep learning (DL)? DL is just a technical subset of machine learning. A particular way of doing machine learning, using connectionist systems or neural networks. The most useful notion to begin to understand the difference is knowing that deep learning is machine learning.
The differences between ML and DL, therefore, do not lie in their use or application, which are the same as those mentioned previously, but in technical issues, such as data consumption, the need for human pre-processing of data — feature engineering — , and, in general, the complexity of the problems dealt with, being DL the most used technique to attack larger data sets.
