Several years ago, I worked on a project where the goal was to try to come up with an "equitable" version of a measure of dependence; the idea was you could take a large multi-dimensional data set, score the dependence for each pair of variables, rank the pairs by their score, and then look at the top-scoring paris to try to determine the most interesting relationship to follow up on in further work. We were motivated by the need for data exploration tools for multi-dimensional data sets.

After a large number of years, we've updated the site http://www.exploredata.net/ , with some (finally) recently published papers, and new versions of the code that are faster, more accurate, and can do additional tasks (what we call TIC as well as MIC). Our technical information subpage has links to papers, including the relatively recent papers in JMLR and the Annals of Applied Statistics. Our MINE-Application page contains links to our new version of the code, as well as links to other versions (such as minepy, a library that has APIs in python and Matlab).

The incentive for all this was, in part, one of the co-authors, Yakir Reshef, finishing up his PhD thesis. Congratulations Yakir!

## Sunday, April 15, 2018

## Wednesday, April 11, 2018

### Sublinear Algorithms Workshop

I was asked to post to announce the workshop/bootcamp on Sublinear Algorithms, June 10-13 at MIT. I plan to be there and possibly talk about some new work.

From the web page (which you should go to to register, if you plan to attend!):

From the web page (which you should go to to register, if you plan to attend!):

### Synopsis

As big data is getting bigger, there is a need for analyzing data with sublinear constraints -- that is, for algorithms which require only sublinear time, space, measurements and/or samples. The goal of this workshop is to bring together experts in various areas of Computer Science, Electrical Engineering, Statistics and Mathematics to discuss recent work and exciting new challenges. It is hoped that the multidisciplinary nature of the workshop will highlight common goals and themes, as well as to facilitate an interchange of technical ideas that may be of use more widely than previously thought. The workshop will be preceded by a one day bootcamp on June 10, with the goal of presenting the basic techniques, definitions and goals in several of the communities.## Monday, March 26, 2018

### An Ad-Hoc Committee on Sexual Harassment and Related Issues

The following is from Yuval Rabani, regarding a joint initiative we are moving forward with to establish policies, procedures, and institutions to deal with harassment and related ethical issues. You may see the post on other theory-related blogs as well.

------------

Recently, many theoreticians have become aware of issues, stories, and rumors concerning sexual harassment within our community, in other CS communities, and more broadly in science.

A number of initiatives, most notably the mushrooming codes of conduct at theory conferences, are already being put into practice.

In consultation among some of the main organizations running theory venues (IEEE TCMF/FOCS, ACM SIGACT/STOC+JACM, EATCS/ICALP, SIAM/SODA+SICOMP) we’ve decided to appoint a joint committee to discuss and propose coordinated policies, procedures, and institutions to deal with harassment and related ethical issues which cut across organizational boundaries. Sandy Irani will chair the committee. Its charter is stated as follows:

"We are setting an ad-hoc committee to draft a proposal for joint ToC measures to combat discrimination, harassment, bullying, and retaliation, and all matters of ethics that might relate to that. Proposed measures may include, but are not restricted to, coordinating policies and guidelines, and setting community-wide institutions for reporting and oversight. The primary goal should be a determination to deter and root out such behavior in the theory community. The issues of false reporting and due process should be taken into account. The committee is expected to conduct the necessary research on existing practices. The committee will submit a report to the appointing organizations by September 30, 2018.”

If you wish an organization be included in the loop, please contact me. If you wish to convey to the committee ideas and thoughts, please contact Sandy or other members as they’ll be announced.

In the meantime, while we are waiting for the committee’s more thoughtful suggestions, here are a couple of simple and potentially effective steps, off the top of my head:

1. If you are harassing someone, please stop.

2. If you are not harassing anyone, please don’t start.

I will gladly contribute to a lively open discussion and react to comments, especially if they occasionally reach my awareness by relaying their existence to my email feed. (Regrettably, I don’t spend all my waking hours monitoring theory blogs.)

## Thursday, March 22, 2018

### Swedish Summer School

I was asked to post the following notice for the upcoming Swedish Summer School for (theoretical) computer scientists. I gave some lectures for it a couple of summers back, and really enjoyed it. Maybe the students did also.

A warning that Djuronaset is not in the city of Stockholm, but a bit over an hour away from downtown Stockholm by bus. While not in the city, it's a beautiful area for walks, and the facilities are very nice. (In particular, they have wonderful saunas in the hotel that should be used daily.)

----------------------------------

A warning that Djuronaset is not in the city of Stockholm, but a bit over an hour away from downtown Stockholm by bus. While not in the city, it's a beautiful area for walks, and the facilities are very nice. (In particular, they have wonderful saunas in the hotel that should be used daily.)

----------------------------------

The 5th Swedish Summer School in Computer Science (http://s3cs.eecs.kth.se/) will be held August 5-11, 2018, in the beautiful Stockholm archipelago at Djuronaset (http://djuronaset.com/en/). The school runs for a full week Monday-Friday in early August when Sweden is at its loveliest, with arrival on Sunday evening and departure Saturdaymorning.

We will celebrate our 5th anniversary by going significantly out of our comfort zone and learn about quantum computation. Ronald de Wolf (https://homepages.cwi.nl/~rdewolf/ ) will give a series of lectures accessible to people who do not know quantum from before. The idea is that this will help all those of us who routinely skip intimidating quantum talks at workshops and conference to overcome our deepest fears and learn enough so that during the next conference we can confidently go to the quantum sessions and actually understand some of what is going on (and maybe even ask a smart question or two).

Another reason for being scared about quantum is that the crypto systems we know and love might no longer be safe. But fear not: Oded Regev (https://cims.nyu.edu/~regev/) will give lectures on lattices and cryptography, explaining, among other things, how to survive in a post-quantum world. Other exciting topics that Oded will touch upon are the Learning with errors (LWE) problem and fully homomorphic encryption.

The summer school is primarily intended for PhD students, but postdocs and bright MSc students are also warmly welcome (and also faculty, subject to availability of slots).

The application deadline is April 20, 2018. Please see http://s3cs.eecs.kth.se/ for more information including instructions for how to apply. Any questions can be directed to s3cs-2018@kth.se.

## Tuesday, March 06, 2018

### Optimizing Learned Bloom Filters with Sandwiching

For the small-ish subset of people out there who care about "learned Bloom filters" (the subject of my last post), I have a small-ish update. I guess the data structure has been unconsciously sitting around in the back of my head, and when I paged it back into my conscious self, I noticed there was an (in hindsight) obvious improvement.

Recall that the learned Bloom filter uses a (magic) black-box predictor obtained from learning theory that predicts whether an element is in a set or not; to avoid false negatives, it then uses a backup Bloom filter to rescue any set elements that have been incorrectly predicted as not being in the set. This leads to two sources of false positives, from the predictor and the backup Bloom filter.

It turns out it is better to get rid of some of those false positives up front. That is, you get better performance if you have a regular old Bloom filter up front, followed by the learned predictor, followed by a (much smaller) backup Bloom filter to round up stray false negatives. Because you have the predictor sandwiched between two Bloom filters, I refer to this as the "sandwiched learned Bloom filter".

The math is surprisingly easily -- and at the same time, really beautiful in terms of the result. It turns out that, if you think of distributing "Bloom filter bits" out to the up-front Bloom filter and the backup Bloom filter, the right thing to do is first put bits on the backup Bloom filter, up to a certain

**fixed amount**, which depends on the the parameters of the predictor (false positive rate, false negative rate). After that, all the bits should go to the up-front Bloom filter.

Like most Bloom filter optimizations, the gains are worthwhile but do not seem to be another-factor-of-two; it looks like it this can cut the size another 10-25% or so, depending on the settings, and it essentially free.

The whole writeup is less than 2 pages, so rather than write and explain it all here, if you're interested, check the arxiv draft. Hopefully, more fun in this setting will be forthcoming....

## Thursday, January 25, 2018

### Some Notes on "Learned Bloom Filters"

About a month ago, a draft paper was put out on arxiv called The Case for Learned Indexed Structures by the people at Google Brain. The paper has received some significant attention, perhaps because of its pedigree and perhaps because it makes some rather bold claims. For example, a line from the abstract states "In this exploratory research
paper, we start from this premise and posit that all existing index structures can be replaced with other
types of models, including deep-learning models, which we term learned indexes."

There has already been some pushback on some of the claims from the draft paper (see here for a response to the paper's discussion of B-trees, and here for a discussion regarding that cuckoo hash tables seem to do just as well as "learned" hash tables). I also read the paper and had some back and forth with the authors about some thoughts and concerns I had; they have been both responsive and thoughtful, and I think the conversation was fruitful on both sides. (At least, they helped me understand their work, and I hope my suggestions for them were useful.) In what follows, I'll explain some of specific thoughts and concerns regarding their proposed learned Bloom filters. I'll try to keep the blog post self-contained, but you may want to read their paper. Also, I've tried to write my ideas down more formally in a short write-up. Here's a link to a draft; this link may age out or disappear if/when I put this on arxiv.

What follows is my description (so any issues are mine, not the authors of the Learned Index paper).

Recall that a Bloom filter approximately represents a set S in small space. One can ask queries of the form: is y an element of S. If y is an element of S, the answer will always be yes (no false negatives), if not there is some chance of false positive. That is, it trades off some amount of error to save space. A key feature is that the probability that any given input y yields a false positive does not depend on y. (Well, it does after the Bloom filter has been instantiated -- some elements give false positives, some don't. But a priori, before you build the Bloom filter using hash functions, if you don't know the hash functions, you don't know what the false positives will be.)

For the learned Bloom filter, suppose we have a very good (black-box) predictor P, which takes an input x and returns P(x), which should be interpreted as an estimate of the probability that x is an element of S. That is, we have a function that can give a very good "guess" if an input x is an element of S. Then we can use P as a "pre-filter" to improve our Bloom filter as follows. When asked if x is an element of S, compute P(x), and simply return that x is in S if P(x) is larger than some suitable threshold z. This may introduce false positives, but it may also introduce false negatives; there may be elements x of S for which P(x) < z. To deal with those, we use a backup filter, which is just a standard Bloom filter that holds the false negatives from the pre-filter using the predictor P. So now we accept that x is an element of S if P(x) is above the threshold, or x obtains a positive answer from the backup filter. The backup filter also introduces false positives.

The authors of the learned index paper suggest that machine learning allows for good predictors P that require small space, thereby improving the Bloom filter in terms of space while achieving the same false positive probability. Note that the predictor does not necessarily have to be that great; even if half the elements of S are false negatives, the backup filter would then be a little bigger than half the size of a standard Bloom filter, so if the predictor P is suitably small, it is plausible that one could get the same performance with smaller space.

I don't know if this idea is new. I don't think I've seen a probabilistic pre-filter for a Bloom filter before; usually a Bloom filter is a pre-filter to some other part of the system. It seems like an interesting idea; certainly, the potential to layer filters is useful, and putting a filter before a Bloom filter is sensible.

A major issue, though, is that the nature of the guarantees one obtains are quite different for learned Bloom filters and standard Bloom filters, which is a subtlety that I worry may not be clear to all readers of the original paper. In particular, it seems in part that confusion might arise because the term "false positive rate" appears to have slightly different meanings in the Bloom filter community and the machine learning community; in the Bloom filter setting, false positive rate generally refers to that false positive probability, which is independent of the query. In the machine learning setting, the false positive rate is an empirical quantity, which depends on the input data and the queries.

An example may be helpful; here's one I used from the writeup. (This is a "thought-experiment" example, not based on actual experiments.) Suppose we are working with numbers in the range [0,1000000), and our learned Bloom filter is supposed to hold a set S that contains a 1000 elements. Half of those elements are random numbers in the range [1000,2000], and half those elements are random over the rest of the range. A good machine learning function might learn a predictor P that says that if x is in the range [1000,2000], then P(x) is "large" -- say around 1/2 -- and otherwise P(x) is small. It could use a threshold z of around 0.4, in which case about half the elements of S are false negatives, and must be stored in the backup filter. This seems like a reasonable outcome for learning.

What is the false positive rate now? Well, it depends on the queries. If all the queries are uniform over the range [1000,2000], the false positive rate would be quite high -- every query to a non-set element would be a false positive. If the queries are uniform over the range [0,1000000), the false positive rate is quite low, since one will rarely get unlucky and ask a query in the range [1000,2000], and otherwise the only false positives will come from the backup Bloom filter. The point is that the false positive probability depends on the distribution of the queries (as well as the predictor). But there may be many applications where the queries are naturally more often for elements that are "similar to" set elements -- that is, they have high P values, like the values in [1000,2000] that are not elements of S -- which may lead to high false positive rates. And if you don't know the query stream in advance, you don't even know what the false positive rate will be.

The response to this seems to be that if you have data about the queries, you can figure out the false positive rate empirically. That is, if you have a test set of queries that you assume is a sample from a distribution over all possible queries, you can determine the false positive rate of that test set, and use it to predict the false positive rate of future queries. My writeup provides basic definitions to formalize this idea, which seems correct. However, this still assumes that "future queries" will come from the same distribution as the test set. Returning to our example, if in our test set queries are uniform over [0,1000000), but then in the future queries change so that they are uniform over a smaller interval [0,100000), that significantly changes the false positive rate of the filter.

My takeaways were the following:

- The idea of adding learning into data structures is certainly an interesting idea.
- But it's still important to try to set up a theoretical framework, to understand what kinds of guarantees can be obtained.
- At the moment, I can see applications where learned Bloom filters might be helpful -- if your data set has significant structure that can be useful to a machine learning function, if you don't have to deal with insertions and deletions frequently, if you have access to query data, and if you have reason to believe the query stream isn't changing (in a distributional sense).
- At the same time, I can see applications where I would be very concerned about the use of learned Bloom filters, such as in a security setting. Also, modern Bloom filter alternatives, such as cuckoo filters (which use less space, and allow insertions/deletions more easily) should be considered and compared when thinking about using learned Bloom filters.
- There's probably more to think about here. In particular, I wonder if there are other places where additional "filter layering" may be useful.

## Tuesday, January 09, 2018

### Double-blind, ALENEX

I wanted to point people to a pair of blog posts (Part 1, Part 2) by Suresh Venkatasubramanian at the geomblog discussing the experience of having the ALENEX conference go double-blind for reviewing this year, from the perspective of him and his co-chair Rasmus Pagh. The high order bits are that they thought it worked fine despite the additional work, and they recommend doing it again in future years.

I am in favor of double-blind reviewing, which is standard for many other conferences in many other areas, but is somehow considered impossible for CS theory. (Suresh does an excellent job addressing why that has been historically, and why the arguments against double-blind reviewing do not really hold up.) The theory community hasn't historically taken "conflicts of interest" issues seriously, as I've written about before (I guess a rather long time ago!). Double-blind reviewing helps with CoI issues, but as Suresh discusses, also deals with the huge implicit bias issues that I think are also a problem in the theory community.

(Except that, really, calling it implicit bias is something of a misnomer -- in many cases it's really explicit bias. As Suresh notes, in providing responses to common arguments:

But how am I supposed to know if the proof is correct if I don't know who the authors are.

Most theory conferences are now comfortable with asking for full proofs. And if the authors don't provide full proofs, and I need to know the authors to determine if the result is believable, isn't that the very definition of bias?This is a real argument commonly presented at PC meetings. I've had people say, "I haven't verified the proofs, but XXX is one of the authors, so I'm sure it's all right.")

I encourage people to go read Suresh's discussion, and I would like to thank Suresh and Rasmus for taking it upon themselves to perform this experiment.

(If you're interested, you can also read the paper by Tomkins et al discussing double-blind reviewing experiences recently at WSDM as well.)

## Thursday, November 16, 2017

### Harvard CS Concentrators Jump Again

During my time as Area Dean for computer science, the number of computer science majors at Harvard more than doubled. Growth has continued, and according to the Crimson, we're now clearly the second largest concentration at Harvard. (It looks like we just barely surpassed Government last year, and now we're clearly larger. Economics is still bigger.)

When you consider that Applied Math is the 4th largest major group, and that's a group that is largely composed of people who do some mix of Math/Computer Science/Economics, I think the numbers actually underestimate the growth in CS at Harvard.

It's been an exciting time since I started at Harvard in 1999. One could see what was coming. Though, sadly, the administration still seems far from having caught up -- we (like many other CS departments) really could use some more resources coming our way. Universities are a bit slow to respond (sometimes for good reasons), and the speed of these changes is, in context, just remarkable. And, I'd have to say, have added to the fun of being a CS professor. Back in 1999, CS was incorrectly perceived to be a backwater at Harvard, small and unnoticed, but where there was a lot of excitement and energy. Fast forward less than two decades, and now we're clearly at the core of Harvard and its future.

When you consider that Applied Math is the 4th largest major group, and that's a group that is largely composed of people who do some mix of Math/Computer Science/Economics, I think the numbers actually underestimate the growth in CS at Harvard.

It's been an exciting time since I started at Harvard in 1999. One could see what was coming. Though, sadly, the administration still seems far from having caught up -- we (like many other CS departments) really could use some more resources coming our way. Universities are a bit slow to respond (sometimes for good reasons), and the speed of these changes is, in context, just remarkable. And, I'd have to say, have added to the fun of being a CS professor. Back in 1999, CS was incorrectly perceived to be a backwater at Harvard, small and unnoticed, but where there was a lot of excitement and energy. Fast forward less than two decades, and now we're clearly at the core of Harvard and its future.

## Thursday, November 09, 2017

### BARC, Copenhagen

A few summers ago, I had an opportunity to visit Copenhagen, and work with Rasmus Pagh for a month. I (and the family) liked it so much we went back the next summer for a hashing workshop. While algorithms and data structure people were already familiar with Denmark's Aarhus and MADALGO, it was clear to me that Copenhagen was on the rise as an algorithmics center.

That's now been formalized, thanks to a huge new grant for algorithms research centered around a collection of fantastic researchers in Copenhagen -- and they're looking for graduate students, post-docs, visitors, and faculty. Rasmus sent me a blurb that I'm posting below, introducing the new Basic Algorithms Research Copenhagen, or BARC.

This is wonderful in so many ways. First, it is always great generally for our community when algorithmic work is recognized and funded. Second, this is a fantastic group of researchers -- for example, obviously I am biased, but Rasmus and Mikkel Thorup, both individually and together, have just been knocking it out of the park over and over again for years in algorithms and data structures (particularly with regard to hashing). Third, for many, this is a marvelous opportunity. Copenhagen is just a wonderful city, and the chance to go and work there with the top people is something many people will enjoy. (And don't we do better work when we're someplace we enjoy being?)

With all that, let me put up the blurb on BARC, and point you to the home page at barc.ku.dk and the Facebook page at http://fb.me/barcdk. (Note there are deadlines coming up for some of the positions.)

That's now been formalized, thanks to a huge new grant for algorithms research centered around a collection of fantastic researchers in Copenhagen -- and they're looking for graduate students, post-docs, visitors, and faculty. Rasmus sent me a blurb that I'm posting below, introducing the new Basic Algorithms Research Copenhagen, or BARC.

This is wonderful in so many ways. First, it is always great generally for our community when algorithmic work is recognized and funded. Second, this is a fantastic group of researchers -- for example, obviously I am biased, but Rasmus and Mikkel Thorup, both individually and together, have just been knocking it out of the park over and over again for years in algorithms and data structures (particularly with regard to hashing). Third, for many, this is a marvelous opportunity. Copenhagen is just a wonderful city, and the chance to go and work there with the top people is something many people will enjoy. (And don't we do better work when we're someplace we enjoy being?)

With all that, let me put up the blurb on BARC, and point you to the home page at barc.ku.dk and the Facebook page at http://fb.me/barcdk. (Note there are deadlines coming up for some of the positions.)

BARC (Basic Algorithms Research Copenhagen) is new center for foundational research in design and analysis of algorithms. The center, which will be inaugurated onDecember 1, is led by Mikkel Thorup. It aims to

*"attract top talent from around the world to an ambitious, creative, collaborative, and fun environment”*in Copenhagen, Denmark. Other core scientists in the center are Stephen Alstrup, Thore Husfeldt, and Rasmus Pagh. The center is part of a larger effort increasing the number of faculty members in algorithms research, and making Copenhagen a hub for algorithms research. The BARC center has received funding of over 6 million USD for the next 6 years from VILLUM Fonden to employ graduate students and post-docs, and to attract a stream of long- and short-term visitors. People interested in BARC should visit the web pages at barc.ku.dk or follow @barcdk on social media — the first assistant/associate professor, post-doc, and PhD student positions are already announced, with deadlines in November and December.
Subscribe to:
Posts (Atom)