Blog: Alexandros Karatzoglou — the brain behind artificial intelligence
Alexandros is one busy man. I’ve meet Alexandros at We Are Developers AI Congress last year. He is one of those guy’s that really understands AI at it’s core. We discussed about people — how people behave when they are predictable and also about magic in some unpredictable scenarios. When it comes to AI we need someone who understand people and who also understands the logic behind machines and it’s behavior. He is one of the guys that when he starts talking everyone in the room starts taking notes. You can find his researches on his blog and in Google Scholars. I really recommend you have a look.
I think recommender systems is great example of human mind like taste for movies for example. Do you think humans are predictable or unpredictable?
They are predictable (smiles). Unfortunately.
It would be more challenging to work if humans were not so predictable?
They are predictable especially if you look at the masses if you take a group of users, humans and you try to predict how this group is going to behave. Groups are actually more predictable than if you take an individual.
If you look at the masses and you compare film taste from people who live in Barcelona, New York, Vienna when you look at the mass — people who like action films are not based on location so much?
What happens is you have… A lot of people like the same things… But then the same people have some kind of special interest also. More or less everybody like the same stuff — but part from that stuff — you also like something that is different than somebody else likes.
So those specifics are interesting?
We call this the long tail. Because if you look at distribution of movies that people watch about 80% of the people watch all the blockbusters, etc.
But then the movies that are less watched — most of the movies have millions of watches on Netflix or some other movie platform and there are some movies that have ten thousands or hundred thousands of watches so this movies is what we call the long tail. And then what happens is that you get a mixture between popular items and long tail items. Everybody like the same — but also there is this little bit special part.
What improvements could deep learning technology bring to human society?
Many things we think are not possible with machine learning — can be possible. What I like particular are generative models. Nowadays this deep learning models are used to generate art or generate designs to create new ideas.
Like for example Zara uses software to create clothes… So this is also probably one of the use cases?
Yes. There is a quite a bit of interest in this area using deep learning for generative design and also in art. Recently there was a painting created with deep learning networks that was auctioned in one of those auction houses for a thousands of dollars.
That’s amazing, so we have a new direction in art now.
It’s always hard to say what benefits would be for society, sometimes they are unexpected. It’s not necessary what you think.
Data quality is a key for training good models — how do you ensure quality of data is good and those not influence your model in bad way?
Data quality has a lot to do with the fact the way the user interacts with your system. So you have to make sure that — when you build a system end to end it’s not only machine learning that it’s important — but what is also important is the human computer interaction part. It’s important to make sure that the way user interacts with your system is meaningful and it’s not confusing — that it creates a clear signal — a clear data signal — that can you use for your algorithm right. For example this e-commerce store that is a confusing website with strange categories and the user clicks and does not find what he wants and all that data that he generates — are clicks that are not necessarily meaningful.
What roles to web browsers play in let’s say online shopping scenarios?
The browser you are using on the machine you are using is playing a specific role. For example if you are using a Mac computer this will be visible on your signature of your browser, because your browser will tell to the web server that costumer is using a Mac Firefox computer. What is well known in this kind of stores — is that Mac users spend more money — because they are buying Mac products, that means that they are not so price sensitive. So might be that model has been trained on that and model might pick items that are more expensive, so the recommendations might be different yes.
So I noticed that if I clear the browser settings my online store shopping starts from blank…
So there are two things here, the second thing is that Amazon puts a cookie on your computer, using that cookie can identify your past sessions.
Based on the cookie, not on the IP for example?
Usually they do it on the cookie, there are other techniques as well, based on the IP and fingerprinting etc, but most common is the cookie. If you clear it the of course the cookie is gone — everything is gone.
On your final slide (at presentation on WAD AI congress) I noticed one point: scalability should be kept in mind. What do you mean by that?
So for example you can create models that are very difficult to train and if you have a lot of data — your model is very difficult to train, then it’s not going to be scalable in the sense: you need very large amount of time to train model and you are not gonna be able to update it very often and it’s going to be old. Scalability for me it’s like a secondary problem in the sense that hardware — because progress in hardware is so fast that scalability becomes a secondary issue. You can now for example read in the media that this company released this new GPO’s and they are twice as fast — the quickest generation — so you can train twice as big models.
But does this effect quality, if you take smaller amount of data, do models give you also good result?
It depends right. I might not necessarily make a big difference. Usually we see if we take two months of data from this e-commerce stores it’s usually enough to build a very good model. If you take three months of data it’s gonna be bit better.
Isn’t also important like for example for recent data to be accurate, not for example historical data, like what user was doing four months ago?
Yes, you have shifts in the data as well so much much older data, might not be so relevant now.
Last question is about films if you cloud share favorite Artificial Intelligence film for our readers?
Deus Ex Machina. (smiles)
Originally published at http://24itworld.wordpress.com on May 12, 2019.