ARE AI, MACHINE LEARNING, AND DEEP LEARNING ALL THE SAME?
In this article, we explain what Artificial Intelligence, Machine Learning, and Deep Learning truly are. Most of us get easily confused about these so-called ‘buzzwords’, and we often think they are all the same. So, are they?
Let’s break down these terms one by one.
AI (Artificial Intelligence) – Machines Imitating Living Things
When Alan Turing first came up with the idea of computers in the 1940s, any machine capable of doing calculation was considered having artificial intelligence. As technology evolved, AI became a more complicated field that refers to machines or applications that mimic animal or human behavior. While people in recent decades do like to define AI with more emphasis on the possession of human cognitive functions (such as the ‘learning’ phase and problem-solving skills), algorithms borrowing concepts from biology (e.g. swarm intelligence, genetic algorithm) are in fact in the AI textbook as well.
Swarm intelligence, which is the collective behavior of a decentralized system (such as the movement of a flock of birds), has been implemented in artificial intelligence for optimization problems. (Source: STRATIO)
In short, artificial intelligence has encompassed a wide spectrum of topics ever since the term was coined; applications ranging from simple script-programmed machine to the generation of natural language can all fall into this large bucket. It can even involve philosophy issues, such as what can be considered as “intelligence” and “thinking”. Of course, to make AI actually usable and not terrifying like the Terminator movie, people have learned to focus on the more applicable topics in AI, such as machine learning and deep learning.
ML (Machine Learning) – An Approach for Better Results
Machine learning is a sub-field in AI in which it mainly uses statistical methods to develop models that can perform certain tasks. That still sounds pretty vague, so we use an example to illustrate.
Imagine you have a set of data containing the studying time of students versus their grades in a test. Assuming there is some connection between the efforts spent on studying and the test score, we can use machine learning model (which normally contains a general structure of mathematical function) to go through the data and try to tune the parameter in the direction that allows the model to produce results closest to the actual data. Since this technique does not rely on human to hard-code specific rules in the model and instead involves model “searching” for the optimal parameters by itself (under a limited function scope), we consider this process as “learning”.
Machine learning can be seen as a way of approximation, as a result it can perform more accurate predictions if given more data. Numerous types of machine learning models were developed in the past decades; these models can be generally classified into supervised and unsupervised learning.
Supervised learning refers to algorithms that learn based on labelled input data. Labelled data can be considered as the ground truth (or answer) for models; during the learning phase, the model aims to minimize the difference between predicted result and the ground truth. So, you can say that the ground truth is “supervising” the model when it is learning. We mostly use supervised learning for regression and classification problems.
Unsupervised learning, on the other hand, is any model trained without using labeled data. With no ground truth, the method can only attempt to model the data with probability densities over the input. While it might not produce results comparable to that of supervised learning, it can still generate useful insight when one does not have the resource to obtain a decent amount of labelled data. Clustering and dimensionality reduction are exemplary applications of unsupervised learning.
If you want to explore more machine learning models, you can take a look at this article to find out more about different types of ML models.
Deep Learning – ML in a more complex form (and so much more computation)
Deep learning is a more advanced branch in machine learning with aims to generate the same regression/classification model with the only difference of using artificial neural networks instead of other statistical models. So…. you can think of DL as a subfield of ML as well. But why is it such a big thing now?
Artificial neural network was (as named) a network of artificial neurons, while an artificial neuron is a mathematical unit that delivers some output value after performing simple calculations on input values. It resembles an actual neuron cell because the latter pretty much does the similar thing: receive electrical signals and produce output. And that’s what an artificial neural network does; it receives data value and allows all the neurons in the network to manipulate with the data value before spitting out some final value. Since such a network requires many neuron layers (stacked after one another) for it to work, we use “deep” to emphasize its spectacular structure.
The general structure of an artificial neural network. (Source: VIASAT)
One might wonder why such simple neural units, after being grouped together in parallel and in series with great quantity, can collectively bring out some of the best applications in the AI industry. It is not so difficult to imagine, in fact; just compare it with the actual course of biological evolution. Single-celled organisms don’t seem to be too functional; but as nature grouped more cells together to make multicellular organisms, more complex life forms emerged. And then invertebrates evolved to vertebrates, fish evolved to primates, cave man evolved to homo sapiens (i.e. us) that built pyramids and shot rockets into space. If you can grasp the power of a group of cells, you can pretty much understand why deep learning is powerful.
For sure there are some catches for all good things; in deep learning, you need to have a large amount of data to be statistically sufficient to train a model (that is why the term “big data” emerged with the surge of deep learning). And to process all these data you need to have powerful hardware to conduct thousands (if not millions) of rounds of computations. Even so, deep learning has already shown us just how much it can achieve. Take a look at Facebook and Pinterest for example; Pinterest uses deep learning to perform better image classification, while Facebook uses similar technology to do facial recognition when you upload your photos. (Remember how Facebook automatically shows little boxes pointing out your friends before you even start to type?)
Apart from social media, deep learning also powers tools that we use in our everyday life, such as Siri, GMail (spam detection), self-driving cars, etc. For engineers, deep learning is considered as a future of AI for it enables seemingly unlimited options of practical ML applications.
Long story short, DL is a part of ML, and ML is a part of the AI field. Thanks to deep learning, our lives have become more convenient. If you want to take advantage of AI implementation for your business to stay competitive, CloudMile can help you with ‘AI service’, the end-to-end AI solution for your business with seamless integration. Make your call to our representatives today!