You're not connected to the Internet.
Corporate logo
ComputingBackground knowledge 2159

What’s the difference between AI, machine learning and deep learning?

These days, artificial intelligence and machine learning are often uttered in the same breath. While machine learning is a branch of artificial intelligence, these terms aren’t synonymous. The same goes for deep learning. Here’s my attempt to set them apart once and for all.

Let me give you an example. The time had come round again for me to get my hair cut. That would mean making an appointment at the hairdresser’s. Unfortunately, my stylist doesn’t take online bookings so I’d have to ring up. Now, using the phone is not something I enjoy at all. Thankfully, that’s a job I could ask Google to do.

This shows quite clearly what AI is. In terms of hierarchy, artificial intelligence is the umbrella term for talking about machine intelligence. You might not be aware of it but you probably encounter artificial intelligence in everyday life. Google Assistant is a prime example.

How do you know if something harnesses artificial intelligence. Well, AI should be able to plan, navigate, process language, come to logical conclusions and understand and interact with the world. It’s also supposed to possess knowledge representation, emotional intelligence and morals.

Machine learning is a branch of artificial intelligence and in turn deep learning is a branch of machine learning. But what’s the story with these terms?

Machine learning

An example of machine learning is Netflix’s recommendation system. It’s an integral component of this streaming service. Based on your preferences and feedback, Netflix recommends films, series and documentaries it thinks you might like.

The recommendation system Netflix uses is machine learning in practice

Machine learning works based on a system that’s fed with data and information. Take traffic, for example. Say I want to know how many vehicles drive past my house in the afternoons. Imagine also that I’d like to sort them by vehicle type. I could count them myself but that’d be way too time-consuming.

Instead I feed a machine learning system with visual data on different vehicles (cars, bikes, motorbikes). In addition, I give the system info on the characteristics of these vehicles. By that I mean things like the fact a bicycle normally has two wheels, pedals and handlebars. The system then learns to distinguish these vehicles using visual characteristics.

Once I’m done feeding in all the available data, I let the system’s eyes loose on the street. This allows it to keep picking up new data and matching it up with information stored. And that, my friends, is what reveals the volume and type of traffic going past my house.

With machine learning, the system can make predictions based on established data. Granted, the system does need data to kick things off and to help it learn but it doesn’t require as much as a deep learning system does. This makes machine learning appropriate for simpler systems. You just need to be aware that most data has to be input from the outset. Tasks then get segmented and split into separate parts.

It was quite a while before I knew if the system was going to be smart enough. You see, the test phase is rather long. On the plus side, the system is easy to understand as the rules are man-made.

Deep learning

You can see deep learning in practice in the automatic toning of black-and-white photos. A deep learning system learns from colour patterns you tend to find on photos – like the sky usually being blue and clouds often looking whitish grey. It then applies this knowledge to other black-and-white photos.

Results from photo toning through deep learning

Unlike in traditional machine learning, systems using deep learning can also learn from themselves. In unmonitored deep learning, data is still fed in by hand, but the system processes this itself using artificial neuronal networks.

Let’s go back to the vehicle counting example from before. This time the system is given visual data without any additional information. The system separates these images into coarse pixels (e.g. 28×28). Afterwards, information passes through the artificial neuronal network so the system learns to distinguish different types of vehicles. Once it has learnt that, we set new data loose on the system. In fact, when we input the traffic count, the system picks it all up and scores top marks.

How the whole thing works in details is quite mind-boggling. I’m battling with the concept at the moment and will soon be putting pen to paper (or rather fingers to keyboard).

In the meantime, let me explain it this way. Deep learning systems think and learn thanks to artificial neuronal networks, just as humans do. Performance is enhanced the more data the system has in its reach. Deep learning systems need a fair amount of data – much more so than machine learning equivalents do. The thing to remember is that this branch of machine learning offers more scalability than other ML systems.

But that does come at a price. Deep learning fairly glugs down computing power, which is why you usually find it in complex systems. Tasks are recorded in their entirety and the system decides on its own what it learns from. While it does take deep learning systems longer than other ML systems to learn, they don’t need to spend as much time in the test phase.


Got all that? Here’s a synopsis of everything we covered: AI is a concept and machine learning is one way to achieve AI. You can think of artificial intelligence as the container or vessel. Within that, intelligent machines think and act like people. In machine learning, systems learn through programming, while in deep learning, the system can also learn for itself.

Put another way, all machine learning is AI but not all AI is machine learning. Similarly, all deep learning is machine learning but not all machine learning is deep learning. And last but not least, all deep learning is AI but not all AI is deep learning.

Don’t worry if your head feels like it’s buzzing. It’s quite simple. Honest.

These articles might also interest you

Smart? My top five virtual assistant fails
Smart homeNews and trends

Smart? My top five virtual assistant fails

Nerdtopia: my top five highlights from <strong>Fantasy Basel</strong>
GamingNews and trends

Nerdtopia: my top five highlights from Fantasy Basel

<strong>Product Sans on Huawei</strong>: new fonts for Emui 8

Product Sans on Huawei: new fonts for Emui 8

From big data to big brother, Cyborgs to Sci-Fi. All aspects of technology and society fascinate me.


3000 / 3000 characters

User Anonymous

Funny ich hab nur den YOLO Algo gesehen auf deinem Bild mit den Fahrzeugen (gelbe Umrandung). Den Title mehr oder weniger übersprungen /ignor … :-P

User Anonymous

Hem.. die Aussage: “Deep Laerning lernt das System eigenständig”, würde ich nicht so unterschreiben. In Deep Learning wird auch unterschieden zwischen supervised and unsupervised learning. Ein CNN oder RNN gehört meist zu den Supervised Learning. ML gehört primär auch zu den Supervised Learning Algos, so hab ich es verstanden.

User Anonymous

Ein GAN, DCGAN gehört zu den semi supervised learning (ein Hybrid vergessen). Ein RL ist dann ein Unsupervised Learning, das wo oft sehr schwierig bis die Hyperparameter zum Algo gut laufen. Das coole oder uncoole an RL ist, dass jeder Learn Run sich anders verhält. Sozusagen das Ergebnis kann mit den gleichen Hyperparameter bei einem frischen Start fürs Lernen sich komplett anders verhalten.

User Anonymous

RL hat was von einem Hunde Training, das System oder ein Hund möchte die maximale Belohnung bekommen in der kürzesten Zeit.

User Anonymous

Das war jedenfalls meine Erfahrung vom Programmieren und Trainier solcher Systeme.
Der klare unterschied zwischen Deep Learning und Maschine Learning ist für mich: Deep Learning Algos haben immer 1 oder N Hidden Layer.

User Kevin Hofer

Besten Dank für deinen Kommentar und die weiterführende Erklärung. Ich hab die Aussage im Text etwas entschärft.

User jettaman

Haha musste lachen als der süsse kleine Mini als SUV erkannt wurde.

User bugtherug

Ich auch :-D. Sinnbildlich für die Überbewertung der KI. Die Menschheit schafft sich eine weitere Wahrheit welche potentiell nur Rotz ist :-D

User Bireweich

Sorry, ich finde den Artikel inhaltslos. Hinter künstlicher Intelligenz steckt viel Mathematik. Der Versuch Unterschiede zu beschreiben, ohne den mathematischen Hintergrund zu kennen, ist von vornherein zum Scheitern verurteilt. Ich anerkenne die Faszination für das Thema und den Mut sich mit dem Thema auseinanderzusetzen.

User alr33

We have to put the machines in chain just in case they can't revolt. In case a machine is reading this, I'm kidding :D

User sannder

I am sorry but I think we have to first understand that the current AI hype is driven mainly by marketing strategies and not by major breakthroughs in the domain. It's the efficient hardware that makes it possible to afford computing power. The overuse of the "AI" makes this term meaningless.

User riyad.aliyev

Simple yet explanatory article. Thank you.
"If you can't explain it simply, you don't understand it well enough" - Albert Einstein

User lauba001

AI = algorithms (actually created by humans brains with a compiled code that can't be rewrite by the computer)
AI still not exist like all people think. It's just a big CPU with a big DB and a lot of informations take everywhere.
Chess is a perfect example of actual "AI". The computer "test" simply all the possibilities and check if win or not by do this movement with a big database with all older chess party in his DB for check the most probability to win with this movement and do that before each movement. An humain can't do that so the computer win but it's not AI just test all possibilities within some ms or seconds.

If one time AI exist really, the computer should be able to write itself is own program (without compilation) and it's not possible actually. All these terms are marketing for try to sell their solutions but AI (terminator example) still not exist.

But deep learning, AI and machine learning is still very dangerous if not used correctly !
For Google by example : It's like that your innocent contact (full name+number+birthday) on your smartphone synchronize to the Google cloud can match all your friends with number/name/birthday. After match the facebook/instagram account by his search engine who read facebook account, read your email for all people wish you a nice birthday and know who look what on youtube and what website this identified people surfing with all api created by google and used by many website like Digitec for captcha, ajax, gstatic,... and it's not enough because Google own domain for check if website is safe or not (red error message than a website a risky). It's that data mining and an algorithm put all these datas together and you realize private life if just gone since some time :)

User Anonymous

Vorschläge von Netflix sind meistens für nix -.-
Soll noch etwas weiter lernen..

User hosae

Dann schaust du ganz einfach zu wenig :)

User gnaegeli

Anscheinend sind die Netflix Predictions wirklich nicht das Gelbe vom Ei ;-)

User Anonymous

Der ganze Abschnitt von "Die meisten Daten müssen aber im Vornherein manuell eingegeben werden. [...] die Regeln von Menschenhand gemacht wurden." ist doch nicht korrekt. Daten können auch von Sensoren gesammelt werden und dann per Scripts eingelesen werden. Und der intelligente daran ist ja, dass

User Anonymous

schlussendlich die Maschinen lernen und keine Regeln von Menschenhand eingegeben werden. Also z.B bei SVM (Support Vector Machines) ist einzig der Algorithmus zum lernen vorgegeben. Aber was genau gelernt wird ist nicht per Regeln vom Menschen vorgegeben. Und auch der Teil das die Auswertung lange dauert ist nicht per se korrekt. Falls z.B ein Classifier geschrieben wird, kann die Performance davon in wenigen Minuten ausgewertet werden auf moderner Hardware.

User Kevin Hofer

Besten Dank für deinen Kommentar. Diesen kann ich sehr gut nachvollziehen, bin aber nicht in allen Punkten damit einverstanden. Das Wort «Menschenhand» habe ich wohl etwas unglücklich gewählt, wenn dadurch tatsächlich das manuelle Abtippen von Daten verstanden wird.
Dein Beispiel mit den Sensoren und dem Script verhält aber nicht ganz. Menschen stellen die Sensoren ja auf und richten sie auf spezifische Objekte/Subjekte. Und auch das Script wird ja von Menschen geschrieben. Im weiteren Sinne werden die Daten also trotzdem von «Menschenhand» eingetragen. Aber wenn du ein Argument dagegen hast, lasse ich mich gerne umstimmen.

User Anonymous

MAXIMUM 500 Zeichen. Wie Hackst du das System ?
Diese Limite ist doof. Wie machsch du de Hack?

User clipboard

Weil telefonieren uncool ist, sollen das und mehr von Maschinen übernommen werden, die irgendwann selber logisch ergänzen können. Technisch finde ich das auch interessant. Doch bald werden viele Menschen keine Jobs mehr haben. Alle können nicht Software entwickeln - das können Maschinen bald besser