Anualmente procuro novos candidatos a orientação nos cursos de Mestrado e Doutorado em Ciência da Computação do Programa de Pós-Graduação em Computação da UFPel. Tenho interesse em projetos na área de Aprendizado de Máquina em específico e Ciência de Dados e Inteligência Artificial em geral. Recentemente, tenho dado enfoque em Aprendizado Profundo e Redes Neurais Convolucionais — tanto aplicações como teoria. Mais em meu site pessoal e ResearchGate.

Candidatos tanto ao Doutorado como Mestrado devem possuir interesse e experiência em pesquisa científica. Capacidade de comunicação em inglês é essencial e fluência é altamente desejável para candidatos ao doutorado. Ao contrário do mestrado, candidatos ao doutorado devem ter possibilidade de se dedicarem em tempo integral ao curso pelo menos a partir do segundo ano de estudos, independente de obterem ou não bolsa. 

Candidatos devem propor projetos de interesse próprio e não procurar algo que pensem que eu me interessaria. No entanto, devem ter intersecção com meus interesses (nota: se não há intersecção com meus interesses, não significa que o projeto deva mudar mas sim que deve-se procurar outro orientador). É fundamental que o candidato possua amplo interesse no projeto sendo proposto e alguma experiência na área do projeto em geral (particularmente para o doutorado). Projetos de doutorado devem demonstrar que o candidato conhece profundamente o estado-da-arte bem como os métodos clássicos para abordar o problema e propor efetivamente algo novo a ser abordado. Projetos de mestrado também devem demonstrar conhecimento na área e estado-da-arte, mas propostas podem ser derivativas, como aplicações de técnicas conhecidas em novos domínios.

Candidatos não devem me procurar objetivando auxílio na escrita de seus projetos. Isto é essencial para garantir isenção na avaliação, uma vez que não tenho como auxiliar todos que me procuram e não desejo beneficiar os mais próximos. No entanto, podem me procurar se estiverem em dúvida se seria um projeto que me interessaria ou que eu teria condições de orientar — e, novamente, em caso negativo isso significa provavelmente que deve-se procurar outro orientador e não que o projeto deva ser alterado!

Misinterpreting Exponential Growth

| No Comments
Imagine that the plot below represents the number of users over time for your favorite web service. Can you find the moment where the service starts to pick up and become successful? 


If you're like most people, you suspect something happened around x=40. Maybe a change in CEO? Maybe a positive review at TechCrunch? Given any service and a plot that looks like this, you'll probably be able to find something to explain a noticeable change in behavior - from a stagnating service to a stellar growthYou may have also thought that the change is from mediocre growth to exponential growth. The term "hockey stick growth" may have come to your mind, if you're familiar with it. 

However, there is nothing different at x=40, or any other point for that matter. The plot merely shows an exponential growth throughout the whole range (y=2^x). The numbers just happen to become large enough at around x=40 for our visual system to start to think there is a change in the underlying model. If you zoomed in at any point of the plot, you would see the same behavior, as in the plot below (showing the same data, but with x from 0 to 25). 


I've seen this mistake happen over and over in different contexts. Exponential curves are very tricky to be analysed by visual inspection and it is really hard not to think something magical happens at some point. But this point is arbitrary - a change in the representation (as zooming in) changes the point - and I'm sure one could find a nice story that fits nicely and "explains" what happened at that point, especially if this story favors whoever is telling it.

So the next time you see a curve like those above, know that whatever was done right (or wrong) was done from the start.

Agents, Memory and Sitcoms

Back in 2009 I published a paper on how agents competing in a bounded-rationality scenario (the Minority Game) could do worse by being able to remember more than their peers. This happened in a very particular evolutionary case, where a majority of agents all had the same memory size and were able to "sync" their behavior, leaving the brainier agent out of sync. 

A reviewer once mentioned this was not realistic and something had to be wrong, since a rational agent receiving more information could decide to filter out the information that was not useful and effectively get in sync with the others if this was beneficial. This was not a central argument in the paper, but I put some thought into this and argued something along the lines that more information meant a larger search space, thus requiring more processing to find reasonable solutions - i.e. information overload. 

I had forgotten about this particular argument until earlier today, when the popular sitcom The Office (US, not the UK version) was not renewed for a new season. The sitcom was in a steep decline after Steve Carrell left (I'll go out on a limb and say that the decline started even earlier) and the announcement was greeted with comments of relieve by several friends. "Finally", they would say, expressing that it should have been cancelled years ago.

You certainly can relate to this behavior in many other situations. A crappy movie based on a popular book is released and fans all regret that the movie was ever made. Or a sequel (aham, Phantom Menace) turns out to be less than what was expected and everyone is thinking they should have left the original trilogy alone. A top actor does an unbelievable bad movie and we all feel he should have quit acting before that. For some reason only entertainment examples come to my mind, but you get the idea.

Psychologically, why do we have this reaction? The fact that there is a Phantom Menace should not reduce the value of Star Wars. Why is it that we can't ignore the latest movies, or the last couple of years of Carrell-less The Office? There is a parallel here with the agents in the Minority Game - we have more information, but we feel we are worse off by knowing more. And if we consider overall satisfaction with a franchise as the utility function, than we really are worse off. We can't choose to ignore what we know. "What has been seen cannot be unseen", goes a popular meme.

This must be related to an experiment mentioned in Daniel Kahneman's fantastic Thinking, Fast and Slow book. In this experiment, volunteers are asked to put a price in a set of items. Another group priced the exact same set added with a few broken items. For someone pricing both sets simultaneously, it wouldn't make sense to price the latter any less than the former. But what happens is that people give the set containing the extra broken items a much lower price. Adding items can reduce the overall price! 

Since price is induced by how much we value something, it is only natural that the same phenomena applies even when no explicit pricing is involved. Adding a few lower-than-average seasons reduce our general satisfaction towards otherwise great sitcoms. It may not be rational, but it is how we work nonetheless.

New site, new blog

It's been several years since I blogged with some regularity and at least a couple of years since I last updated my personal website. So I'm starting a new one and throwing in the package a new blog too (this one). The reason I spent too long without doing both is the same as everyone else's: social network sites. 

Blogs had (have?) trackbacks, comments, blog rolls. All these features added a social layer on top of the blogosphere, but were a pain to set up and maintain. Twitter made too easy to just throw some idea out of the virtual window and build a social network around a profile. Linkedin promised to be the one-stop place for my professional identity, while Facebook would be my personal identity. 

The trouble with outsourcing this sort of information keeping is that you give away some control over information that is yours. That's fine, of course, until it isn't. It is mildly annoying that you can't retrieve your oldest tweets, but Techcrunch is the most recent voice of an increasing chorus raising the concern of letting companies decide what you can and can't say and calling for a (return to a) federated web. That means to have full control of your information, on private servers and your own domain.

All that lead me to decide that I want to take some of the control back. That does not mean that I'm deleting my profiles in every social network site (I might do it for a few... I'm looking at you, Orkut). It is still a pain to keep the social layer going on in blogs, so the idea is to use those services as pointers to information hosted here. Instead of writing a big post in Facebook, I'll just post in here and send an excerpt with a link to Facebook, Twitter, Google+. If any of these services die (and I assume most will), it won't take my information with it.

When you think about it, we sort of do this with personal photos - we keep them in our hard drives, DVDs or whatever kids are using to store GB of photos, and just send small-ish versions of them to social network sites. We don't even consider deleting the local copy after uploading it. But we seem to be okay with doing this with text. 

I'm keeping all the posts of the old blog separate from this one. I just checked, and the first post in there is from 2003, but I have even older ones in some backup around here (I think my very first blog post was around 2000, we even called them by their full name, weblogs). I'm sure I'm not impressing anyone with these numbers. Anyway, this one is a fresh start. Let's see how long it lasts. Oh, and I'll probably switch a lot between English and Portuguese. I find that some thoughts are better expressed in English. Go figure.