Nobody in the world knows how to train one hidden layer

I said it.
In recent years, there has been a lot of buzz about deep learning, where the learning algorithm is not based on the Bayes rule and probability. People are optimizing arbitrary, complicated cost functions, and they are doing it with gradient descent, so they don't even reach the minimum of the (incorrect) function that they want to optimize.
I just wanted to remind that nobody in the world knows how to train one hidden layer well, so perhaps instead of hand-waving about deep learning that much, which gets annoying, it may be worth to examine again simpler, fundamental models.

My publication on the Netflix Prize is now free. Download it here.
The previous 4-page publication has so far over 450 citations, and the newer publication has 195 pages and 0 citations.

So read it, cite it. My h-index is 1, and I want to increase it to 2. I don't feel like a real scientist with an h-index 1.

You have to read it to not stay behind your competition.

My thoughts on AlphaGo

AlphaGo is winning 2-0 with one of the best go players in the world.
(Edit: it won 3-0)
(Edit2: they played all five games. It won 4-1.)

I happen to know well a similar domain of computer chess.
In 2005 my chess playing program won the championship of my country. The unique thing about my program was that I successfully used machine learning to learn the weights of my program's evaluation function. I used value adaptation and move adaptation from games of >2300 ELO players, and I used also learning from self-play. The learning methods used in my program were similar to those used in Logistello and Deep Blue. The difference was that the creators of Deep Blue ultimately used weights chosen by hand, and my program used automatically learned weights.

So I am one of few people in the world, who understands the challenge and benefits of using machine learning in computer chess, go, and similar board games.

My thoughts:
1. I guess that the most recent large improvements of the skill of AlphaGo come from improving leaf evaluation, and moving from Monte Carlo Search and playouts to an algorithm closer to alpha-beta search. Previously they had to use playouts to evaluate leafs, to correct inaccuracies of static leaf evaluation. Playouts are always nonoptimal - they should be replaced with a properly designed search and static evaluation.
2. The goal of the AlphaGo project is PR for Google. I think that they downplay the effort put in feature engineering, and they exaggerate the role of deep learning.
3. Best human go players aren't that good at playing go as advertised.
4. Computer go is not as hard as advertised.

Random Strangers

My new project Random Strangers is a random chat like Chatroulette or Omegle, but combined with a recommendation engine. Users rate conversations on a scale 1-5, and the idea is to match users that will like each other, based on predicted ratings. This should solve the well known Chatroulette problem (I always wanted to solve humanity's biggest problems).

5000 best things

It's time to sum up my growing website - lists of 5000 best things. It started as a simple, searchable list of 5000 movies. The response was enthusiastic (hacker news front page, wykop front page and an article in Gigazine retweeted about 1000 times), so I decided to put extra work into it and explore the opportunity. So far I created four lists: movies, books, websites and videos.

The movies part is the most advanced, with instant search, filtering by 13 genres, film/TV series or by year, links to IMDb, Wikipedia, Rotten Tomatoes, Netflix, etc., 54 different rankings of the 5000 movies, and additionally, one personalized ranking calculated by a recommendation engine (just like in my all previously written recommendation engines, a new ranking is calculated and displayed immediately at the moment of rating a movie). 5000 best websites are divided into 24 categories, and 5000 best videos into 35 categories.

What to do next? All suggestions are welcome. Create more lists or extend the existing ones? Or put some serious effort into marketing? People tell me that I don't advertise the website.

I got 9k visits from stumbleupon in the last weeks to websites/Porn. Makes me think - should I listen to the market?




The blog was renamed to "What I Do", because this is what this blog ultimately became - news about my projects for anyone interested. I stole the subtitle from Remi Gaillard. I am thinking about starting another, more technical blog on programming, machine learning, the Internet and business (I know nothing about business), but who would read that. It kind of does not make sense to write into the void. And another blog about life and everything (I know nothing about life, but I don't think it matters).

I wrote a book

Or rather, I call it a monograph.

Why a book? I wanted to do something for humanity. Still, I am not that concerned with humanity to give away all fruits of my work for free.

The title is "Predicting movie ratings and recommender systems" and you can buy download it here. It's pretty obscure, specialistic stuff on what I understood about the Netflix Prize data, recommender systems, and about prediction tasks in general. Definitely not a publication for everyone.

So now after I have done something for humanity, in the rest of my life I will: 1) earn money and 2) do whatever I want.

Movie discovery and recommendations

As a side effect of digging into the Netflix Prize data I created a set of flash applications. For a long time I was the only user, but I am not that beyond. I finished it up lately and it is ready to share. It has colors and all. Users should like it.

I launched it under the name "The Galaxy of Movies". What's inside: two ways of visualizing similar movies, a 2D recommender system, and a quiz game.

Features: interactive 2d maps, search, filtering by classic and experimental genres, recommendations for two people, option of importing ratings. The application is self-contained, and is smaller than 150k together with all data. The recommendations are built-in, calculated within the flash application, without having a specialized server.




I named the six most contributing features (hidden genres) learned by the regularized SVD, creating a set of 12 new, experimental genres (after discretizing) - they can be used to filter the movies:
* Realism vs. Idealization
* Safety vs. Surprise
* Fairy Tale vs. Distrust
* Feminism vs. Testosterone
* Innocence vs. Heroism
* Growing up vs. Journey

From the point of view of rating prediction those new genres carry much more information (allow to better assess if you will like the movie) than the standard genres, like Comedy, Drama, etc.

Moving on with one thing

A difficult subject that I feel I have to explain. I do not want to repeat myself to everyone, so I will just write it once here. In 2004 I enrolled in "PhD studies" in computer science at Warsaw University. It turned out that the Institute of Informatics not only has zero didactic offer for people like me, but over all these years I did not manage to convince the university that what I do is worth any funding or any other kind of support. So I was close to this decision every year. I do not keep in touch with those people since long time, so this is rather stating the obvious - I do not want to have a PhD from that place.

The whole situation is odd, because my publication, with over 170 citations, is the most cited on the whole faculty since 2007. My understanding of all that, is that my presence bothered much those various eternal beneficiaries of the system. Maybe it's I am too stubborn, and in such places more welcome are people who bow their heads low. Well, this is how the system works and nothing can be done about it.

Predictions for soccer

4th place in Kaggle's World Cup "confidence challenge". This time all predictions were "by hand", using the bookmakers' odds.

Netflix Grand Prize & what's new

1. The Netflix Prize is over, won by the 7-person team BellKor's Pragmatic Chaos. Congratulations to the winners.

I ended up in 43rd place individually, and in 34th place in the two-person team with John Tomfohr.
My best reached place was 3rd in June 2007, ten months after the start of the contest, and the best reached place of our team was 2nd, at the moment of creating it, one day before the first Progress Prize in October 2007.

2. arek-paterek.com is now my homepage. Basically, that website will be the center of my presence on the web, with my current e-mail, links to things I share, and so on.

3. Some summary of my service lolrate/svdsystem "recommender system for everything". I launched it in December 2008, it was online for a few months, and had about 300 users. Now closed. It will probably evolve into another project.

What's new 2

Back in Poland. Time to get back to work on the PhD.

What's new

With John Tomfohr we gave a talk at Morgan Stanley in New York about our Netflix Prize solutions.

Some time later we met again at Stanford Startup School. I liked these two talks most: [1] [2]

US, Santa Cruz and chess

I moved for some time to the US. Since one month I'm settling in in Santa Cruz, California, visiting UCSC.

I wouldn't be myself without looking for a place to play chess. It turned out that every week there is a tournament in a cafe in a large bookstore in the middle of Santa Cruz downtown. I played there and later I was surprised to find on the Internet an analysis of one of my games. I lost a game to Dana Mackenzie, and you can follow the analysis on his blog. Dana is a freelance writer, former mathematics teacher, and holds PhD from Princeton. His USCF rating happens to be equal to my FIDE rating, 2048.

Sharpgame.net

I launched a new service - a simple online game:
sharpgame.net - English version
gra-w-karteczki.net - Polish version

Netflix Prize: Progress Prize

Team Belkor/Korbell from AT&T Research won the $50,000 Progress Prize (see links to the press coverage). Congratulations!
According to the rules, they were required to describe their solution - this is a very interesting paper.

I've joined John Tomfohr in team basho and we are currently in 4th place with the score 8.25%. The leaders have the score 8.50%.

Netflix Prize: Ideas giveaway

I learned a lot today. A very short report from the conference KDD 2007, San Jose, California:
My paper, poster and slides.
The workshop proceedings.

Netflix Prize: The Long Walk

ML@UToronto, a team representing one of the strongest machine learning research groups, regains 3rd place with their new solution.

Netflix Prize: Even more progress

3rd place in the Netflix Prize.

Netflix Prize: More progress

I am in 4th place in the Netflix Prize with the score over 7%.
This is one of those rare times when I can say I am quite pleased with myself.

Netflix Prize: Progress

My last submission is in ninth place. Team Bellkor takes the lead.

Netflix Prize: The truth is out there, in the data

Nine months after the start of the contest, still in 10th place, less than 1% behind the leading team.

Currently leading is a four-person team Gravity from Budapest University of Technology and Economics. In third place is a four-person team from University of Toronto, including Prof. Geoffrey Hinton, a well known person in the machine learning community.

Looks like the competitors approach the task really professionally. A recent article in the New York Times, "Netflix Prize Still Awaits a Movie Seer", cites the leader of Team Gravity:
"Domonkos Tikk, a data mining expert who is a senior researcher at the university in Budapest, leads Team Gravity. Dr. Tikk said that since October, his team, which is composed of three Ph.D. candidates and himself, has spent eight hours a day, seven days a week on the problem".