From reingold at stanford.edu Mon May 1 12:43:48 2017
From: reingold at stanford.edu (Omer Reingold)
Date: Mon, 1 May 2017 12:43:48 -0700
Subject: [theory-seminar] Group photo during theory lunch
Message-ID:
Dear friends,
This Thursday at 12 we will be meeting in front of bytes (opposite to
Gates) for a group photo (which hector is gracious enough to take). After
the photo we will continue to theory lunch. If you feel part of our theory
community then (regardless of your affiliation), please join us for a photo
(which will be posted online).
See you Thursday,
Omer
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
From ashwinpp at stanford.edu Mon May 1 14:19:24 2017
From: ashwinpp at stanford.edu (Ashwin Pradeep Paranjape)
Date: Mon, 1 May 2017 21:19:24 +0000
Subject: [theory-seminar] Algorithms reading group
Message-ID:
Hi all,
We'll have our first reading group this Wednesday (5/3) 3-5 pm in the theory lounge.
Zhou Fan shall be presenting his own work "How Well Do Local Algorithms Solve Semidefinite Programs?"(https://arxiv.org/abs/1610.05350) from STOC 17.
Abstract
--
Several probabilistic models from high-dimensional statistics and machine learning reveal an intriguing --and yet poorly understood-- dichotomy. Either simple local algorithms succeed in estimating the object of interest, or even sophisticated semi-definite programming (SDP) relaxations fail.
In order to explore this phenomenon, we study a classical SDP relaxation of the minimum graph bisection problem, when applied to Erd\H{o}s-Renyi random graphs with bounded average degree d>1, and obtain several types of results. First, we use a dual witness construction (using the so-called non-backtracking matrix of the graph) to upper bound the SDP value. Second, we prove that a simple local algorithm approximately solves the SDP to within a factor 2d2/(2d2+d?1) of the upper bound. In particular, the local algorithm is at most 8/9 suboptimal, and 1+O(1/d) suboptimal for large degree.
We then analyze a more sophisticated local algorithm, which aggregates information according to the harmonic measure on the limiting Galton-Watson (GW) tree. The resulting lower bound is expressed in terms of the conductance of the GW tree and matches surprisingly well the empirically determined SDP values on large-scale Erd\H{o}s-Renyi graphs.
We finally consider the planted partition model. In this case, purely local algorithms are known to fail, but they do succeed if a small amount of side information is available. Our results imply quantitative bounds on the threshold for partial recovery using SDP in this model.
Hope to see you there.
Thanks
Ashwin
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
From michael.kim at cs.stanford.edu Wed May 3 10:06:57 2017
From: michael.kim at cs.stanford.edu (Michael Kim)
Date: Wed, 3 May 2017 10:06:57 -0700
Subject: [theory-seminar] Fwd: Group photo during theory lunch
In-Reply-To:
References:
Message-ID:
Reminder! At the start of Theory Lunch tomorrow (noon), we will meet *in
front of Bytes* (Packard Building) for a group photo. Then, lunch will be
in Gates 463 as usual. Omer will be leading the program for theory lunch
with a discussion about kicking off the Theory Dish blog. Hope to see you
all there for the photo and lunch!
---------- Forwarded message ----------
From: Omer Reingold
Date: Mon, May 1, 2017 at 12:43 PM
Subject: [theory-seminar] Group photo during theory lunch
To: theory-seminar at lists.stanford.edu, Hector Garcia-Molina <
hector at cs.stanford.edu>
Dear friends,
This Thursday at 12 we will be meeting in front of bytes (opposite to
Gates) for a group photo (which hector is gracious enough to take). After
the photo we will continue to theory lunch. If you feel part of our theory
community then (regardless of your affiliation), please join us for a photo
(which will be posted online).
See you Thursday,
Omer
_______________________________________________
theory-seminar mailing list
theory-seminar at lists.stanford.edu
https://mailman.stanford.edu/mailman/listinfo/theory-seminar
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
From reingold at stanford.edu Wed May 3 10:25:50 2017
From: reingold at stanford.edu (Omer Reingold)
Date: Wed, 03 May 2017 17:25:50 +0000
Subject: [theory-seminar] Fwd: Group photo during theory lunch
In-Reply-To:
References:
Message-ID:
And let me also emphasize that regardless of whether you plan to take part
in the blog or not and which is your affiliation, we would love you join
for the photo. If you feel part of our theory community then your place is
there :-)
Omer
On Wed, May 3, 2017 at 10:07 AM Michael Kim
wrote:
> Reminder! At the start of Theory Lunch tomorrow (noon), we will meet *in
> front of Bytes* (Packard Building) for a group photo. Then, lunch will
> be in Gates 463 as usual. Omer will be leading the program for theory
> lunch with a discussion about kicking off the Theory Dish blog. Hope to
> see you all there for the photo and lunch!
>
> ---------- Forwarded message ----------
> From: Omer Reingold
> Date: Mon, May 1, 2017 at 12:43 PM
> Subject: [theory-seminar] Group photo during theory lunch
> To: theory-seminar at lists.stanford.edu, Hector Garcia-Molina <
> hector at cs.stanford.edu>
>
>
> Dear friends,
>
> This Thursday at 12 we will be meeting in front of bytes (opposite to
> Gates) for a group photo (which hector is gracious enough to take). After
> the photo we will continue to theory lunch. If you feel part of our theory
> community then (regardless of your affiliation), please join us for a photo
> (which will be posted online).
>
> See you Thursday,
> Omer
>
> _______________________________________________
> theory-seminar mailing list
> theory-seminar at lists.stanford.edu
> https://mailman.stanford.edu/mailman/listinfo/theory-seminar
>
>
> _______________________________________________
> theory-seminar mailing list
> theory-seminar at lists.stanford.edu
> https://mailman.stanford.edu/mailman/listinfo/theory-seminar
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
From moses at cs.stanford.edu Wed May 3 10:26:58 2017
From: moses at cs.stanford.edu (Moses Charikar)
Date: Wed, 3 May 2017 10:26:58 -0700
Subject: [theory-seminar] theory day at Google, May 12
Message-ID:
Folks,
If you are on the TOCA-SV mailing list, you got this already ... If not,
here is program and registration info for the theory day at Google on May
12 (next Friday):
Registration Link (takes less than a minute to register and helps us
prepare visitor badges, etc.): https://goo.gl/rnpxWB
Information Page: https://sites.google.com/view/toca-sv-day-may-2017/home
I encourage you all to attend. If transport is a concern, that shouldn't
stop you from registering. We will coordinate transport for everyone who is
planning to go from Stanford.
We're still looking for people to sign up for short talks. Sign up here:
https://docs.google.com/document/d/1HBAuV_Zxf8k6XrR1-6czxKdrvupK-LKSnONrJDpIrgM/edit
Cheers,
Moses
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
From michael.kim at cs.stanford.edu Thu May 4 12:08:07 2017
From: michael.kim at cs.stanford.edu (Michael Kim)
Date: Thu, 4 May 2017 12:08:07 -0700
Subject: [theory-seminar] Fwd: Group photo during theory lunch
In-Reply-To:
References:
Message-ID:
We are in front of the tree in front of Huang
On May 3, 2017 10:26 AM, "Omer Reingold" wrote:
> And let me also emphasize that regardless of whether you plan to take part
> in the blog or not and which is your affiliation, we would love you join
> for the photo. If you feel part of our theory community then your place is
> there :-)
>
> Omer
>
> On Wed, May 3, 2017 at 10:07 AM Michael Kim
> wrote:
>
>> Reminder! At the start of Theory Lunch tomorrow (noon), we will meet *in
>> front of Bytes* (Packard Building) for a group photo. Then, lunch will
>> be in Gates 463 as usual. Omer will be leading the program for theory
>> lunch with a discussion about kicking off the Theory Dish blog. Hope to
>> see you all there for the photo and lunch!
>>
>> ---------- Forwarded message ----------
>> From: Omer Reingold
>> Date: Mon, May 1, 2017 at 12:43 PM
>> Subject: [theory-seminar] Group photo during theory lunch
>> To: theory-seminar at lists.stanford.edu, Hector Garcia-Molina <
>> hector at cs.stanford.edu>
>>
>>
>> Dear friends,
>>
>> This Thursday at 12 we will be meeting in front of bytes (opposite to
>> Gates) for a group photo (which hector is gracious enough to take). After
>> the photo we will continue to theory lunch. If you feel part of our theory
>> community then (regardless of your affiliation), please join us for a photo
>> (which will be posted online).
>>
>> See you Thursday,
>> Omer
>>
>> _______________________________________________
>> theory-seminar mailing list
>> theory-seminar at lists.stanford.edu
>> https://mailman.stanford.edu/mailman/listinfo/theory-seminar
>>
>>
>> _______________________________________________
>> theory-seminar mailing list
>> theory-seminar at lists.stanford.edu
>> https://mailman.stanford.edu/mailman/listinfo/theory-seminar
>>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
From moses at cs.stanford.edu Thu May 4 15:01:39 2017
From: moses at cs.stanford.edu (Moses Charikar)
Date: Thu, 4 May 2017 15:01:39 -0700
Subject: [theory-seminar] talks today
Message-ID:
There are at least a couple of talks today late afternoon that folks on
this list might find interesting:
--Moses
Math Department Colloquium: 4:30pm in Math 380-W
Yufei Zhao, Oxford University
"Large deviations in discrete random structures"
Abstract:
What is the probability that the number of triangles in an Erd?s?R?nyi
random graph exceeds its mean by a constant factor? This problem has been a
useful litmus test for concentration bound techniques. Even the order of
the log-probability was considered a difficult problem until its resolution
a few years ago by Chatterjee, and independently, DeMarco and Kahn. We now
wish to determine the exponential rate of the tail probability. I will
highlight some recent methods and results on this problem and its variants.
Thanks to recent developments by Chatterjee, Varadhan, Dembo, and Eldan,
this large deviations problem reduces to a natural variational problem,
which then can be solved (in some instances) to deduce the asymptotics of
the log-probability. I will explain this large deviation principle and its
consequences. No background knowledge will be assumed.
------------------------------------------------------------
------------------------------------------
ISL Colloquium, Packard 101 Time: 4:15 - 5:15 pm
Title: Geometries of Word Embeddings
Speaker: Professor Pramod Viswanath, University of Illinois at
Urbana-Champaign
Abstract:
Real-valued word vectors have transformed NLP applications; popular
examples are word2vec and GloVe, recognized for their ability to capture
linguistic regularities via simple geometrical operations. In this talk, we
demonstrate further striking geometrical properties of the word vectors.
First, we show that a very simple, and yet counter-intuitive,
postprocessing technique, which makes the vectors "more isotropic",
renders off-the-shelf vectors even stronger. Second, we show that a
sentence containing a target word is well represented by a low-rank
subspace; subspaces associated with a particular sense of the target word
tend to intersect with a line (one-dimensional subspace). We harness this
Grassmannian geometry to disambiguate (in an unsupervised way) multiple
senses of words, specifically so on the most promiscuously polysemous of
all words: prepositions. A surprising finding is that rare senses,
including idiomatic/sarcastic/metaphorical usages, are efficiently
captured. Our algorithms are all unsupervised and rely on no linguistic
resources; we validate them by presenting new state-of-the-art results on a
variety of multilingual benchmark datasets.
References:
1. Geometry of Compositionality, AAAI '17, https://arxiv.org/abs/1611.09799
2. Geometry of Polysemy, ICLR, '17, https://arxiv.org/abs/1610.07569
3. Representing Sentences as Low-rank subspaces, ACL '17,
https://arxiv.org/abs/1704.05358
4. Prepositions in Context, preprint, https://arxiv.org/abs/1702.01466
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
From whkong at stanford.edu Mon May 8 11:16:31 2017
From: whkong at stanford.edu (Weihao Kong)
Date: Mon, 8 May 2017 18:16:31 +0000
Subject: [theory-seminar] Theory Seminar-Dan Alistarh
Message-ID:
Hi Friends,
Dan Alistarh will be speaking in the Theory Seminar this week, May 11th Thursday 4:15pm in Gates 463A. See you there.
Data Structures of the Future: Concurrent, Optimistic, and Relaxed
A central computing trend over the last decade has been the need to process increasingly larger amounts of data as efficiently as possible. This development is challenging both software and hardware design, and is altering the way data structures are constructed, implemented, and deployed.
In this talk, I will present examples of such new data structure design ideas and implementations. In particular, I will discuss some inherent limitations of parallelizing classic data structures, and then focus on approaches to circumvent these limitations. The first approach is to relax the software semantics, to allow for approximation, randomization, or both. The second is to modify the underlying hardware architecture to unlock more parallelism. Time permitting, I will also cover results showing that both approaches can improve real-world performance, and touch upon some of the major open questions in the area.
Short bio: Dan Alistarh is an Assistant Professor at IST Austria, currently visiting ETH Zurich on an SNF Ambizione Fellowship. Previously, he was a Researcher at Microsoft Research, Cambridge, UK, and a Postdoctoral Associate at MIT CSAIL. He received his PhD from the EPFL, under the guidance of Prof. Rachid Guerraoui. His research focuses on distributed algorithms and concurrent data structures, and spans from algorithms and lower bounds to practical implementations.
???????????????
Nika Haghtalab will be the speaker next week. Title/Abstract to be announced.
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
From moses at cs.stanford.edu Tue May 9 00:16:52 2017
From: moses at cs.stanford.edu (Moses Charikar)
Date: Tue, 9 May 2017 00:16:52 -0700
Subject: [theory-seminar] carpooling for Friday theory event at Google
Message-ID:
Folks,
We hope to see many of you at the TOCA-SV event this Friday.
https://sites.google.com/view/toca-sv-day-may-2017/home
If you plan to drive and can take passengers, or if you are looking for a
ride, please coordinate carpooling via this shared Google doc:
https://docs.google.com/document/d/1_DvZQ1LFCsG4aDW8kmNEfat2XrFBK2qHt2c7SLLrZ38/edit?usp=sharing
Please register if you haven't done so already:
https://goo.gl/rnpxWB
Cheers,
Moses
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
From reingold at stanford.edu Wed May 10 13:40:14 2017
From: reingold at stanford.edu (Omer Reingold)
Date: Wed, 10 May 2017 13:40:14 -0700
Subject: [theory-seminar] Theory Dish blog is live
Message-ID:
Hi Everyone,
Our blog https://theorydish.blog/ is now live and on blog aggregator. A 2nd
post is coming soon.
Whether you plan to write posts, take part in discussions through the
comment section or even just read - I hope you will all view it as *your *blog
and *ours*. Having said that, please do consider writing your first post
(and don't hesitate asking for my help and the help of your colleagues).
Yalla Balagan*,
Omer
* A combination of Hebrew and Arabic meaning - lets go and make some
noise/mess
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
From michael.kim at cs.stanford.edu Wed May 10 15:20:21 2017
From: michael.kim at cs.stanford.edu (Michael Kim)
Date: Wed, 10 May 2017 15:20:21 -0700
Subject: [theory-seminar] Theory Lunch 5/11 - Luna Frank-Fischer
Message-ID:
Hi all,
Join us for theory lunch tomorrow at noon in Gates 463. Luna will be our
speaker tomorrow -- see her title and abstract below. Hope to see you
there!
******
Speaker: Luna Frank-Fischer
Title: Locality via Partially Lifted Codes (or, do you even lift, bro?)
Abstract: In error-correcting codes, locality refers to several different
ways of quantifying how easily a small amount of information can be
recovered from encoded data. In this work, we study a notion of locality
called the *s*-Disjoint-Repair-Group Property (*s*-DRGP). This notion can
interpolate between two very different settings in coding theory: that of
Locally Correctable Codes (LCCs) when *s* is large?a very strong
guarantee?and Locally Recoverable Codes (LRCs) when *s* is small?a
relatively weaker guarantee. This motivates the study of the *s*-DRGP for
intermediate *s*, which is the focus of our paper. We construct codes in
this parameter regime which have a higher rate than previously known codes.
Our construction is based on a novel variant of the lifted codes of Guo,
Kopparty and Sudan. Beyond the results on the *s*-DRGP, we hope that our
construction is of independent interest, and will find uses elsewhere.
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
From moses at cs.stanford.edu Wed May 10 17:18:21 2017
From: moses at cs.stanford.edu (Moses Charikar)
Date: Wed, 10 May 2017 17:18:21 -0700
Subject: [theory-seminar] STOC 2017 student travel grants
Message-ID:
Just FYI ... if you are planning to attend STOC 2017, tomorrow is the
deadline for applying for student travel grants:
http://acm-stoc.org/stoc2017/travel-support.html
--Moses
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
From moses at cs.stanford.edu Thu May 11 08:25:30 2017
From: moses at cs.stanford.edu (Moses Charikar)
Date: Thu, 11 May 2017 08:25:30 -0700
Subject: [theory-seminar] Ron Graham today, 3pm
Message-ID:
Ron Graham is speaking at the combinatorics seminar at 3pm today. (details
below)
At 4:15pm, we have our theory seminar (Dan Alistarh on Data Structures of
the Future).
Cheers,
Moses
Combinatorics Seminar
3pm ? 4pm, Building 380, room 384-H
The combinatorics of solving linear equations
Ron Graham (UCSD)
Description
One of the fundamental topics in combinatorics involves deciding whether
some given linear equation has solutions with all its variables lying in
some restricted set, and if so, estimating how many such solutions there
are. In this talk, we will describe some of the old and new results in this
area, as well as discuss a number of unsolved problems.
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
From whkong at stanford.edu Thu May 11 13:32:53 2017
From: whkong at stanford.edu (Weihao Kong)
Date: Thu, 11 May 2017 20:32:53 +0000
Subject: [theory-seminar] Theory Seminar-Dan Alistarh
In-Reply-To:
References:
Message-ID: <0B319BAE-D0DF-48FC-ACC1-E1AD17C069B6@stanford.edu>
As a reminder, we will have Theory Seminar 4:15pm today at Gates 463A.
On May 8, 2017, at 11:16 AM, Weihao Kong > wrote:
Hi Friends,
Dan Alistarh will be speaking in the Theory Seminar this week, May 11th Thursday 4:15pm in Gates 463A. See you there.
Data Structures of the Future: Concurrent, Optimistic, and Relaxed
A central computing trend over the last decade has been the need to process increasingly larger amounts of data as efficiently as possible. This development is challenging both software and hardware design, and is altering the way data structures are constructed, implemented, and deployed.
In this talk, I will present examples of such new data structure design ideas and implementations. In particular, I will discuss some inherent limitations of parallelizing classic data structures, and then focus on approaches to circumvent these limitations. The first approach is to relax the software semantics, to allow for approximation, randomization, or both. The second is to modify the underlying hardware architecture to unlock more parallelism. Time permitting, I will also cover results showing that both approaches can improve real-world performance, and touch upon some of the major open questions in the area.
Short bio: Dan Alistarh is an Assistant Professor at IST Austria, currently visiting ETH Zurich on an SNF Ambizione Fellowship. Previously, he was a Researcher at Microsoft Research, Cambridge, UK, and a Postdoctoral Associate at MIT CSAIL. He received his PhD from the EPFL, under the guidance of Prof. Rachid Guerraoui. His research focuses on distributed algorithms and concurrent data structures, and spans from algorithms and lower bounds to practical implementations.
???????????????
Nika Haghtalab will be the speaker next week. Title/Abstract to be announced.
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
From whkong at stanford.edu Thu May 11 16:15:12 2017
From: whkong at stanford.edu (Weihao Kong)
Date: Thu, 11 May 2017 23:15:12 +0000
Subject: [theory-seminar] Theory Seminar-Dan Alistarh
In-Reply-To: <0B319BAE-D0DF-48FC-ACC1-E1AD17C069B6@stanford.edu>
References: ,
<0B319BAE-D0DF-48FC-ACC1-E1AD17C069B6@stanford.edu>
Message-ID: <354EF53A-F786-4E8E-9806-44450120BB06@stanford.edu>
Now!
Sent from my iPhone
On May 11, 2017, at 1:33 PM, Weihao Kong > wrote:
As a reminder, we will have Theory Seminar 4:15pm today at Gates 463A.
On May 8, 2017, at 11:16 AM, Weihao Kong > wrote:
Hi Friends,
Dan Alistarh will be speaking in the Theory Seminar this week, May 11th Thursday 4:15pm in Gates 463A. See you there.
Data Structures of the Future: Concurrent, Optimistic, and Relaxed
A central computing trend over the last decade has been the need to process increasingly larger amounts of data as efficiently as possible. This development is challenging both software and hardware design, and is altering the way data structures are constructed, implemented, and deployed.
In this talk, I will present examples of such new data structure design ideas and implementations. In particular, I will discuss some inherent limitations of parallelizing classic data structures, and then focus on approaches to circumvent these limitations. The first approach is to relax the software semantics, to allow for approximation, randomization, or both. The second is to modify the underlying hardware architecture to unlock more parallelism. Time permitting, I will also cover results showing that both approaches can improve real-world performance, and touch upon some of the major open questions in the area.
Short bio: Dan Alistarh is an Assistant Professor at IST Austria, currently visiting ETH Zurich on an SNF Ambizione Fellowship. Previously, he was a Researcher at Microsoft Research, Cambridge, UK, and a Postdoctoral Associate at MIT CSAIL. He received his PhD from the EPFL, under the guidance of Prof. Rachid Guerraoui. His research focuses on distributed algorithms and concurrent data structures, and spans from algorithms and lower bounds to practical implementations.
???????????????
Nika Haghtalab will be the speaker next week. Title/Abstract to be announced.
_______________________________________________
theory-seminar mailing list
theory-seminar at lists.stanford.edu
https://mailman.stanford.edu/mailman/listinfo/theory-seminar
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
From whkong at stanford.edu Mon May 15 16:08:31 2017
From: whkong at stanford.edu (Weihao Kong)
Date: Mon, 15 May 2017 23:08:31 +0000
Subject: [theory-seminar] Two Theory Seminar this week
Message-ID: <4ED7DFB7-DDC4-4046-B4C2-1C338EBE438D@stanford.edu>
Hi Friends,
We will have two Theory Seminar this week, both on Thursday May 18th, first one by Sam Hopkins at 3pm and second one by Nika Haghtalab at 4:15pm. Both seminars will be held on Gates 463A. See you there!
May 18, 2017
Gates 463A, 4:15PM
Nika Haghtalab (CMU)
Oracle-efficient Online Learning and Applications to Auction Design
We consider the fundamental problem of learning from expert advice, a.k.a online no-regret learning, where we have access to an offline optimization oracle that can be used to compute, in constant time, the best performing expert at any point in time. We consider the design of no-regret algorithms that are computationally efficient using such an oracle. We present structural properties under which we show oracle-efficient no-regret algorithms exist, even when the set of experts is exponentially large in a natural representation of the problem. Our algorithm is a generalization of the Follow-The-Perturbed-Leader algorithm of Kalai and Vempala that at every step follows the best-performing expert subject to some perturbations. Our design uses a shared source of randomness across all experts that can be efficiently implemented by using an oracle on a random modification of the history of the play at every time step.
Our second main contribution is showing that the structural properties required for our oracle-efficient online algorithm are present in a large class problems. As an example, we discuss applications of our oracle-efficient learning results to the adaptive optimization of a large class of auctions, including (1) VCG auctions with bidder-specific reserves in single-parameter settings, (2) envy-free item pricing in multi-item auctions, and (3) Myerson-like auctions for single-item settings.
Bio: Nika Haghtalab is a Ph.D. student at the Computer Science department of Carnegie Mellon University, co-advised by Avrim Blum and Ariel Procaccia. Her research interests include machine learning theory, computational aspects of economics, and algorithms. Nika is a recipient of the IBM and Microsoft Research Ph.D. fellowships.
________________________________
May 18, 2017
Gates 463A, 3PM
Sam Hopkins (Cornell)
Sample-optimal inference, the method of moments, and community detection
We propose a simple and efficient meta-algorithm for Bayesian estimation problems (i.e. hidden variable, latent variable, or planted problems). Our algorithm uses low-degree polynomials together with new and highly robust tensor decomposition methods. We focus on the question: for a given estimation problem, precisely how many samples (up to low-order additive terms) do polynomial-time algorithms require to obtain good estimates of hidden variables? Our meta-algorithm is broadly applicable, and achieves statistical or conjectured computational sample-complexity thresholds for many well-studied problems, including many for which previous algorithms were highly problem-specific.
As a running example we employ the stochastic block model -- a widely studied family of random graph models which contain latent community structure. We recover and unify the proofs of the best-known sample complexity bounds for the partial recovery problem in this model. We also give the first provable guarantees for partial recovery of community structure in constant-degree graphs where nodes may participate in many communities simultaneously. This model is known to exhibit a sharp sample complexity threshold -- with fewer than a very specific number of samples, recovering community structure becomes impossible. While previous explanations for this phenomenon appeal to sophisticated ideas from statistical mechanics, we give a new and simple explanation based on properties of low-degree polynomials.
Joint work with David Steurer.
???????????????
Kent Quadrud and Pravesh Kothari will be speaking next week!
Schedule here: http://theory.stanford.edu/~aflb/2016-17.html
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
From michael.kim at cs.stanford.edu Wed May 17 10:09:51 2017
From: michael.kim at cs.stanford.edu (Michael Kim)
Date: Wed, 17 May 2017 10:09:51 -0700
Subject: [theory-seminar] Theory Lunch 5/18 -- Edith Cohen
Message-ID:
Hi all,
Please join us for theory lunch tomorrow at noon in Gates 463. Edith Cohen
will visit us from Google. Look forward to seeing you there!
******
Speaker: Edith Cohen (Google Research and Tel Aviv University)
Title: Multi-objective weighted sampling and applications to summary
statistics and clustering
Abstract: For a set of keys X and associated values f(x), a weighted
sample by f allows us to estimate the sum of f(x) for query segments of keys
with statistically-guaranteed quality. When we sample with weights f,
however, the guarantees do not hold for estimating sums with respect to
another function g(x). Multi-objective samples are powerful and versatile
summaries that allow us to obtain estimates with statistical guarantees
with respect to multiple f's.
We discuss properties of these samples and showcase two important
application domains.
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
From yuhch123 at cs.stanford.edu Wed May 17 15:20:27 2017
From: yuhch123 at cs.stanford.edu (Huacheng Yu)
Date: Wed, 17 May 2017 22:20:27 +0000
Subject: [theory-seminar] University Oral Exam: Huacheng Yu, Wednesday May 31,
10am, Gates 104
Message-ID:
University Oral Examination
Title: Techniques for Proving Data Structure Lower Bounds
Huacheng Yu
Computer Science Department
Stanford University
Advised by Ryan Williams and Omer Reingold
Wednesday, May 31, 10am, Gates 104
Abstract: We present several techniques for proving the nonexistence of
"too-
efficient" data structures. Then we apply these techniques to prove lower
bounds for three data structure problems: dynamic interval union, 2D
weighted
orthogonal range counting, and 2D parity orthogonal range counting.
In this talk, I will present the lower bound results, and focus on the lower
bound for the 2D parity orthogonal range counting problem. In this problem,
the data structure maintains a set of points S in the 2D space, supporting:
- insertion: add a point to set S;
- query: return the parity of number of points in an axis-parallel
rectangle.
We show that any data structure for this problem with polylog n insertion
time
must have >~ log^1.5 n query time. The lower bound is proved using a new
weak-
simulation theorem. Roughly speaking, if the problem admits an efficient
data
structure, then it is possible to succinctly encode a sequence of updates
(insertions) such that one may answer a query significantly better than
random
guessing based only on the encoding. The lower bound then follows from
showing
that the data structure problem does not have this property.
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
From whkong at stanford.edu Thu May 18 12:00:15 2017
From: whkong at stanford.edu (Weihao Kong)
Date: Thu, 18 May 2017 19:00:15 +0000
Subject: [theory-seminar] Two Theory Seminar this week
In-Reply-To: <4ED7DFB7-DDC4-4046-B4C2-1C338EBE438D@stanford.edu>
References: <4ED7DFB7-DDC4-4046-B4C2-1C338EBE438D@stanford.edu>
Message-ID:
As a reminder, we will start at 3 today!
On May 15, 2017, at 4:08 PM, Weihao Kong > wrote:
Hi Friends,
We will have two Theory Seminar this week, both on Thursday May 18th, first one by Sam Hopkins at 3pm and second one by Nika Haghtalab at 4:15pm. Both seminars will be held on Gates 463A. See you there!
May 18, 2017
Gates 463A, 4:15PM
Nika Haghtalab (CMU)
Oracle-efficient Online Learning and Applications to Auction Design
We consider the fundamental problem of learning from expert advice, a.k.a online no-regret learning, where we have access to an offline optimization oracle that can be used to compute, in constant time, the best performing expert at any point in time. We consider the design of no-regret algorithms that are computationally efficient using such an oracle. We present structural properties under which we show oracle-efficient no-regret algorithms exist, even when the set of experts is exponentially large in a natural representation of the problem. Our algorithm is a generalization of the Follow-The-Perturbed-Leader algorithm of Kalai and Vempala that at every step follows the best-performing expert subject to some perturbations. Our design uses a shared source of randomness across all experts that can be efficiently implemented by using an oracle on a random modification of the history of the play at every time step.
Our second main contribution is showing that the structural properties required for our oracle-efficient online algorithm are present in a large class problems. As an example, we discuss applications of our oracle-efficient learning results to the adaptive optimization of a large class of auctions, including (1) VCG auctions with bidder-specific reserves in single-parameter settings, (2) envy-free item pricing in multi-item auctions, and (3) Myerson-like auctions for single-item settings.
Bio: Nika Haghtalab is a Ph.D. student at the Computer Science department of Carnegie Mellon University, co-advised by Avrim Blum and Ariel Procaccia. Her research interests include machine learning theory, computational aspects of economics, and algorithms. Nika is a recipient of the IBM and Microsoft Research Ph.D. fellowships.
________________________________
May 18, 2017
Gates 463A, 3PM
Sam Hopkins (Cornell)
Sample-optimal inference, the method of moments, and community detection
We propose a simple and efficient meta-algorithm for Bayesian estimation problems (i.e. hidden variable, latent variable, or planted problems). Our algorithm uses low-degree polynomials together with new and highly robust tensor decomposition methods. We focus on the question: for a given estimation problem, precisely how many samples (up to low-order additive terms) do polynomial-time algorithms require to obtain good estimates of hidden variables? Our meta-algorithm is broadly applicable, and achieves statistical or conjectured computational sample-complexity thresholds for many well-studied problems, including many for which previous algorithms were highly problem-specific.
As a running example we employ the stochastic block model -- a widely studied family of random graph models which contain latent community structure. We recover and unify the proofs of the best-known sample complexity bounds for the partial recovery problem in this model. We also give the first provable guarantees for partial recovery of community structure in constant-degree graphs where nodes may participate in many communities simultaneously. This model is known to exhibit a sharp sample complexity threshold -- with fewer than a very specific number of samples, recovering community structure becomes impossible. While previous explanations for this phenomenon appeal to sophisticated ideas from statistical mechanics, we give a new and simple explanation based on properties of low-degree polynomials.
Joint work with David Steurer.
???????????????
Kent Quadrud and Pravesh Kothari will be speaking next week!
Schedule here: http://theory.stanford.edu/~aflb/2016-17.html
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
From moses at cs.stanford.edu Fri May 19 13:16:49 2017
From: moses at cs.stanford.edu (Moses Charikar)
Date: Fri, 19 May 2017 13:16:49 -0700
Subject: [theory-seminar] tensor decomposition guest lecture,
1:30pm-4:20 today
Message-ID:
Just wanted to send this out to the theory and ML folks ...
if you're looking to decompress after the NIPS deadline, Sam Hopkins is
giving a guest lecture in my class today 1:30-4:20 on tensor decomposition
via Sum-of-Squares. Talk will be adaptive to the interests of the
audience. Its in 200-217 (History corner - 5-10 min walk from Gates).
http://web.stanford.edu/class/cs369h/
Cheers,
Moses
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
From jvondrak at gmail.com Mon May 22 11:03:15 2017
From: jvondrak at gmail.com (Jan Vondrak)
Date: Mon, 22 May 2017 11:03:15 -0700
Subject: [theory-seminar] Probability seminar today
Message-ID:
Hi theory folks,
I'm giving a talk in the probability seminar today (Sequoia Hall, room 200,
4:30 - 5:30 pm), which is quite algorithmic and related to the Lovasz local
lemma.
-- Jan
=================================
Title: Computing the independence polynomial with negative arguments by
correlation decay
Abstract:
The work of Shearer, Scott and Sokal established an important connection
between the multivariate independence polynomial and the Lovasz Local Lemma
(LLL). In particular, the maximal root-free region in the negative
orthant containing the origin (the "Shearer region") determines the region
of applicability of the LLL. Motivated by this connection, we study the
question of computing (approximately) the independence polynomial with
negative arguments (in contrast to the positive-argument regime studied
previously by Weitz, Sly and Sun.)
Our main result is that Weitz's algorithm based on decay
of correlations works in the Shearer region, and provides an FPTAS for the
multivariate independence polynomial on bounded-degree graphs (with
quasi-polynomial
running time in general). Together with a recent hardness result
of Galanis-Goldberg-Stefankovic (and an independent work of Patel-Regts),
this provides a complete picture in terms of the complexity of
computing the independence polynomial on the real line. Further, we study
the quantitative question of correlation decay close to the boundary of the
Shearer region, with applications for the LLL: A subexponential algorithm
to decide whether every collection of events with a given dependency graph
has a non-empty complement, and a new deterministic algorithm to find
objects guaranteed by the LLL.
Joint work with Nick Harvey and Piyush Srivastava.
===========================
https://statistics.stanford.edu/events/probability-seminar
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
From whkong at stanford.edu Mon May 22 17:53:13 2017
From: whkong at stanford.edu (Weihao Kong)
Date: Tue, 23 May 2017 00:53:13 +0000
Subject: [theory-seminar] Theory Seminar this week---Pravesh Kothari and
Kent Quadrud
Message-ID:
Hi Friends,
This week we will again have two theory seminars. Pravesh Kothari will be speaking on Wednesday 4:15PM and Kent Quadrud will be speaking on Thursday 4:15PM. Both talks will be held in Gates 463A. See you all there!
May 24, 2017
Gates 463A, 4:15PM
Pravesh Kothari (Princeton)
Quantum Entanglement, Sum-of-Squares, and the Log-Rank Conjecture
In this talk, I will show a exp(?(n)polylogn)exp?((n)polylog?n)-time algorithm for the Best Separable State (BSS) problem (see next paragraph for a classical algorithmic formulation): given an n2?n2n2?n2 quantum measurement matrix MM, distinguish between the cases that there is a separable (i.e., non-entangled) state ?? that M accepts with probability 11, and the case that every separable state is accepted with probability at most 0.99. Aside from being a central question in quantum information theory arising in the study of entanglement, recent works have uncovered potentially useful connections between BSS and fundamental problems in classical algorithm design such as Max-Cut, Small-Set-Expansion, and Unique Games.
Equivalently, the algorithm takes input a subspace W?Rn2W?Rn2 and distinguishes between the case that WW contains a rank one matrix and the case that every rank one matrix is at least ?? far (in the Euclidean distance) from WW.
At the heart of our result is a general rounding paradigm that uses polynomial reweightings to round the solution to the Sum-of-Squares (SoS) semidefinite programming relaxations. Somewhat surprisingly, the construction of such a polynomial reweighting scheme uses an argument inspired by the recent breakthrough on log-rank conjecture by Lovett (STOC?14, JACM?16) who showed that the communication complexity of every rank-n Boolean matrix is bounded by ?npolylog(n)npolylog(n).
I will assume no specialized background in the talk.
Based on joint work with Boaz Barak (Harvard) and David Steurer (Cornell, IAS).
May 25, 2017
Gates 463A, 4:15PM
Kent Quadrud (UIUC)
Near-Linear Time Approximation Schemes for some Implicit Fractional Packing Problems
We consider several implicit fractional packing problems and obtain faster implementations of approximation schemes by a general framework based on multiplicative-weight updates. This leads to new algorithms with near-linear running times for some fundamental problems in optimization. We highlight three concrete applications. The first is to find the maximum fractional packing of spanning trees in a capacitated graph; we obtain a (1??)(1??)-approximation in ~O(m/?2)O~(m/?2) time, where mm is the number of edges in the graph. Second, we consider the LP relaxation of the weighted unsplittable flow problem on a path and obtain a (1??)(1??)-approximation in ~O(n/?2)O~(n/?2) time, where nn is the number of demands. Third, given an undirected edge weighted graph on mm edges, we obtain a (1+?)(1+?)-approximation to the Held-Karp bound for the Metric-TSP instance induced by the shortest path metric in the graph in ~O(m/?2)O~(m/?2) time. Each of these algorithms give an order of magnitude improvement over previous results.
This is joint work with Chandra Chekuri.
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
From michael.kim at cs.stanford.edu Tue May 23 08:56:06 2017
From: michael.kim at cs.stanford.edu (Michael Kim)
Date: Tue, 23 May 2017 08:56:06 -0700
Subject: [theory-seminar] Talk tomorrow
In-Reply-To:
References:
Message-ID:
Hi all,
Greg Bodwin is in town and will give a talk tomorrow at 11am in the
security lab. Hope to see many of you there!
Speaker: Greg Bodwin (MIT)
Title: Optimal Vertex Fault Tolerant Spanners
Abstract:
A stretch-k spanner of a graph G is a sparse subgraph H that preserves the
distances of G up to a multiplicative error of k. The tradeoffs between
the stretch k and the optimal sparsity of a spanner H are very well
understood. However, in many applications, one has to worry about
"failures" in the graph: perhaps a small set of nodes F will break, and
then we still want H \ F to be a good spanner of G \ F. This motivates the
study of "fault-tolerant spanners," where H is an f-fault tolerant spanner
of G if H \ F is a spanner of G \ F for all possible sets F of f nodes.
We close the central open problem in fault tolerant spanners by
establishing that O_k( n^{1 + 1/k} f^{1 - 1/k} ) edges always suffice for
an f-fault tolerant k-spanner, and that this is optimal (under standard
conjectures). This is the first result in a large body of work on fault
tolerant graph compression that achieves a sublinear dependence on f.
Additionally, we are able to avoid considerable complexities in past work
on fault tolerant spanners by constructing our optimal spanners with a very
simple and natural greedy algorithm, which is the obvious extension of the
textbook greedy algorithm used in the non-faulty setting.
Joint work with Michael Dinitz, Merav Parter, and Virginia Vassilevska
Williams.
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
From michael.kim at cs.stanford.edu Wed May 24 10:39:37 2017
From: michael.kim at cs.stanford.edu (Michael Kim)
Date: Wed, 24 May 2017 10:39:37 -0700
Subject: [theory-seminar] Theory Lunch 5/25 -- Vatsal Sharan
Message-ID:
Hi all,
Theory lunch as usual, tomorrow at noon in Gates 463. Vatsal will be our
speaker -- title and abstract below. Hope to see you there!
******
Speaker: Vatsal Sharan
Title: Prediction with a Short Memory
Abstract:
We consider the problem of predicting the next observation given a sequence
of past observations. We show that for any distribution over observations,
if the mutual information between past observations and future observations
is upper bounded by $I$, then a simple Markov model over the most recent
$I/\epsilon$ observations can obtain KL error $\epsilon$ with respect to
the optimal predictor with access to the entire past. We also show that the
sample complexity of the Markov model is nearly optimal for any polynomial
time algorithm based on plausible computational complexity conjectures.
This is based on joint work (https://arxiv.org/pdf/1612.02526v1.pdf) with
Sham Kakade, Percy Liang and Greg Valiant.
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
From michael.kim at cs.stanford.edu Wed May 24 10:42:57 2017
From: michael.kim at cs.stanford.edu (Michael Kim)
Date: Wed, 24 May 2017 10:42:57 -0700
Subject: [theory-seminar] Talk tomorrow
In-Reply-To:
References:
Message-ID:
Reminder: talk in 20 mins!
On Tue, May 23, 2017 at 8:56 AM, Michael Kim
wrote:
> Hi all,
>
> Greg Bodwin is in town and will give a talk tomorrow at 11am in the
> security lab. Hope to see many of you there!
>
> Speaker: Greg Bodwin (MIT)
> Title: Optimal Vertex Fault Tolerant Spanners
>
> Abstract:
>
> A stretch-k spanner of a graph G is a sparse subgraph H that preserves the
> distances of G up to a multiplicative error of k. The tradeoffs between
> the stretch k and the optimal sparsity of a spanner H are very well
> understood. However, in many applications, one has to worry about
> "failures" in the graph: perhaps a small set of nodes F will break, and
> then we still want H \ F to be a good spanner of G \ F. This motivates the
> study of "fault-tolerant spanners," where H is an f-fault tolerant spanner
> of G if H \ F is a spanner of G \ F for all possible sets F of f nodes.
>
> We close the central open problem in fault tolerant spanners by
> establishing that O_k( n^{1 + 1/k} f^{1 - 1/k} ) edges always suffice for
> an f-fault tolerant k-spanner, and that this is optimal (under standard
> conjectures). This is the first result in a large body of work on fault
> tolerant graph compression that achieves a sublinear dependence on f.
> Additionally, we are able to avoid considerable complexities in past work
> on fault tolerant spanners by constructing our optimal spanners with a very
> simple and natural greedy algorithm, which is the obvious extension of the
> textbook greedy algorithm used in the non-faulty setting.
>
> Joint work with Michael Dinitz, Merav Parter, and Virginia Vassilevska
> Williams.
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
From whkong at stanford.edu Wed May 24 11:45:13 2017
From: whkong at stanford.edu (Weihao Kong)
Date: Wed, 24 May 2017 18:45:13 +0000
Subject: [theory-seminar] Theory Seminar this week---Pravesh Kothari and
Kent Quadrud
In-Reply-To:
References:
Message-ID:
Reminder, Pravesh Kothari will be speaking today at 4:15PM.
On May 22, 2017, at 5:53 PM, Weihao Kong > wrote:
Hi Friends,
This week we will again have two theory seminars. Pravesh Kothari will be speaking on Wednesday 4:15PM and Kent Quadrud will be speaking on Thursday 4:15PM. Both talks will be held in Gates 463A. See you all there!
May 24, 2017
Gates 463A, 4:15PM
Pravesh Kothari (Princeton)
Quantum Entanglement, Sum-of-Squares, and the Log-Rank Conjecture
In this talk, I will show a exp(?(n)polylogn)exp?((n)polylog?n)-time algorithm for the Best Separable State (BSS) problem (see next paragraph for a classical algorithmic formulation): given an n2?n2n2?n2 quantum measurement matrix MM, distinguish between the cases that there is a separable (i.e., non-entangled) state ?? that M accepts with probability 11, and the case that every separable state is accepted with probability at most 0.99. Aside from being a central question in quantum information theory arising in the study of entanglement, recent works have uncovered potentially useful connections between BSS and fundamental problems in classical algorithm design such as Max-Cut, Small-Set-Expansion, and Unique Games.
Equivalently, the algorithm takes input a subspace W?Rn2W?Rn2 and distinguishes between the case that WW contains a rank one matrix and the case that every rank one matrix is at least ?? far (in the Euclidean distance) from WW.
At the heart of our result is a general rounding paradigm that uses polynomial reweightings to round the solution to the Sum-of-Squares (SoS) semidefinite programming relaxations. Somewhat surprisingly, the construction of such a polynomial reweighting scheme uses an argument inspired by the recent breakthrough on log-rank conjecture by Lovett (STOC?14, JACM?16) who showed that the communication complexity of every rank-n Boolean matrix is bounded by ?npolylog(n)npolylog(n).
I will assume no specialized background in the talk.
Based on joint work with Boaz Barak (Harvard) and David Steurer (Cornell, IAS).
May 25, 2017
Gates 463A, 4:15PM
Kent Quadrud (UIUC)
Near-Linear Time Approximation Schemes for some Implicit Fractional Packing Problems
We consider several implicit fractional packing problems and obtain faster implementations of approximation schemes by a general framework based on multiplicative-weight updates. This leads to new algorithms with near-linear running times for some fundamental problems in optimization. We highlight three concrete applications. The first is to find the maximum fractional packing of spanning trees in a capacitated graph; we obtain a (1??)(1??)-approximation in ~O(m/?2)O~(m/?2) time, where mm is the number of edges in the graph. Second, we consider the LP relaxation of the weighted unsplittable flow problem on a path and obtain a (1??)(1??)-approximation in ~O(n/?2)O~(n/?2) time, where nn is the number of demands. Third, given an undirected edge weighted graph on mm edges, we obtain a (1+?)(1+?)-approximation to the Held-Karp bound for the Metric-TSP instance induced by the shortest path metric in the graph in ~O(m/?2)O~(m/?2) time. Each of these algorithms give an order of magnitude improvement over previous results.
This is joint work with Chandra Chekuri.
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
From whkong at stanford.edu Thu May 25 10:18:09 2017
From: whkong at stanford.edu (Weihao Kong)
Date: Thu, 25 May 2017 17:18:09 +0000
Subject: [theory-seminar] Theory Seminar this week---Pravesh Kothari and
Kent Quadrud
In-Reply-To:
References:
Message-ID: <6AF0F130-EAF9-4DBB-B9AD-F48C1839E510@stanford.edu>
As a reminder, Kent Quanrud will be speaking today at 4:15 in Gates 463A.
On May 22, 2017, at 5:53 PM, Weihao Kong > wrote:
Hi Friends,
This week we will again have two theory seminars. Pravesh Kothari will be speaking on Wednesday 4:15PM and Kent Quadrud will be speaking on Thursday 4:15PM. Both talks will be held in Gates 463A. See you all there!
May 24, 2017
Gates 463A, 4:15PM
Pravesh Kothari (Princeton)
Quantum Entanglement, Sum-of-Squares, and the Log-Rank Conjecture
In this talk, I will show a exp(?(n)polylogn)exp?((n)polylog?n)-time algorithm for the Best Separable State (BSS) problem (see next paragraph for a classical algorithmic formulation): given an n2?n2n2?n2 quantum measurement matrix MM, distinguish between the cases that there is a separable (i.e., non-entangled) state ?? that M accepts with probability 11, and the case that every separable state is accepted with probability at most 0.99. Aside from being a central question in quantum information theory arising in the study of entanglement, recent works have uncovered potentially useful connections between BSS and fundamental problems in classical algorithm design such as Max-Cut, Small-Set-Expansion, and Unique Games.
Equivalently, the algorithm takes input a subspace W?Rn2W?Rn2 and distinguishes between the case that WW contains a rank one matrix and the case that every rank one matrix is at least ?? far (in the Euclidean distance) from WW.
At the heart of our result is a general rounding paradigm that uses polynomial reweightings to round the solution to the Sum-of-Squares (SoS) semidefinite programming relaxations. Somewhat surprisingly, the construction of such a polynomial reweighting scheme uses an argument inspired by the recent breakthrough on log-rank conjecture by Lovett (STOC?14, JACM?16) who showed that the communication complexity of every rank-n Boolean matrix is bounded by ?npolylog(n)npolylog(n).
I will assume no specialized background in the talk.
Based on joint work with Boaz Barak (Harvard) and David Steurer (Cornell, IAS).
May 25, 2017
Gates 463A, 4:15PM
Kent Quadrud (UIUC)
Near-Linear Time Approximation Schemes for some Implicit Fractional Packing Problems
We consider several implicit fractional packing problems and obtain faster implementations of approximation schemes by a general framework based on multiplicative-weight updates. This leads to new algorithms with near-linear running times for some fundamental problems in optimization. We highlight three concrete applications. The first is to find the maximum fractional packing of spanning trees in a capacitated graph; we obtain a (1??)(1??)-approximation in ~O(m/?2)O~(m/?2) time, where mm is the number of edges in the graph. Second, we consider the LP relaxation of the weighted unsplittable flow problem on a path and obtain a (1??)(1??)-approximation in ~O(n/?2)O~(n/?2) time, where nn is the number of demands. Third, given an undirected edge weighted graph on mm edges, we obtain a (1+?)(1+?)-approximation to the Held-Karp bound for the Metric-TSP instance induced by the shortest path metric in the graph in ~O(m/?2)O~(m/?2) time. Each of these algorithms give an order of magnitude improvement over previous results.
This is joint work with Chandra Chekuri.
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
From moses at cs.stanford.edu Thu May 25 22:21:54 2017
From: moses at cs.stanford.edu (Moses Charikar)
Date: Thu, 25 May 2017 22:21:54 -0700
Subject: [theory-seminar] Statistical vs Complexity Tradeoffs via SoS,
tomorrow 1:30-4:20
Message-ID:
Hi all,
Pravesh Kothari is giving a guest lecture in my class tomorrow: 1:30-4:20
in Gates 463A. Some of you on this list might be interested (see details
below). You are welcome to attend.
Cheers,
Moses
*Statistical vs Computational Complexity Tradeoffs via Sum-of-Squares (SoS)*
Fundamental problems in unsupervised machine learning and
average-case-complexity are concerned with recovering some hidden *signal* or
structure surrounded by random *noise*. Examples include Sparse PCA,
Learning Mixtures of Gaussians, Planted Clique, and Refuting Random
Constraint Satisfaction Problems (CSPs).
The fundamental parameter of interest is the *signal to noise ratio* (SNR)
- the relative strength of signal compared to the noise. Curiously, for
many such problems, there?s a wide gap between the *computational threshold* -
the SNR above which the best known efficient algorithms can recover the
hidden structure and the *statistical threshold* - where the hidden
structure is information theoretically identifiable (using a potentially
non-efficient algorithm).
Are these gaps inherent? Can we predict the computational threshold with
high confidence?
It turns out that the standard machinery of NP-hardness and reductions for
*worst-case* problems is inadequate to understand such trade-offs.
In this context, the goal of this class is threefold:
1) Show that (SoS) method is powerful enough to capture known algorithms
and yield non-trivial new ones. Here, our example would be Refuting Random
CSPs.
2) Show that we can also understand the SoS method well enough to prove
strong lower bounds - we will see the example of Random k-XOR and time
permitting, a sketch of the techniques involved for Random CSPs in general.
We will also see a sketch of the lower bound for planted clique.
3) Hint at the general principle of using the SoS-threshold of the SNR as a
fairly reliable proxy for the computational threshold for some problems.
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
From abboud at CS.Stanford.EDU Fri May 26 19:49:05 2017
From: abboud at CS.Stanford.EDU (abboud at CS.Stanford.EDU)
Date: Fri, 26 May 2017 19:49:05 -0700
Subject: [theory-seminar] University Oral Exam: Amir Abboud, Tuesday May 30,
**1:30PM**
Message-ID: <6baebff670b4229b6a3190fff8aeb5d9.squirrel@xenon.stanford.edu>
University Oral Examination
Title: Hardness in P
Amir Abboud
Computer Science Department, Stanford University
Advised by Virginia Vassilevska Williams and Omer Reingold
When: Tuesday May 30, 2017 1:30PM ***Note time change***
Where: Gates 463A
Abstract:
The celebrated theory of NP-Hardness classifies problems into ?easy? (in
P) and ?NP-hard? (probably not in P). However, while the easy problems do
have polynomial time algorithms, in many cases, a polynomial runtime like
O(n^2) or O(n^3) can be too slow. Many fundamental "easy" problems have
resisted decades of attempts at getting truly fast algorithms, and are
typically solved by potentially inaccurate but fast algorithms in
practice. Such problems include Sequence Alignment, Parsing, Distance
Computation in Graphs, and so many other problems we encounter in basic
Algorithms courses. Can we formally explain the lack of faster algorithms
and justify the use of heuristics, even for ?easy? problems?
In this talk, I will present a theoretical framework for proving hardness
for problems in P. Inspired by NP-Hardness, our approach is based on
reductions which have uncovered fascinating structure among the different
problems in P. We can now point at a small set of core problems, and say:
For dozens of important problems in P there is no hope to get faster
algorithms, until we know how to solve the core problems faster.
From amir.abboud at gmail.com Tue May 30 02:10:32 2017
From: amir.abboud at gmail.com (Amir Abboud)
Date: Tue, 30 May 2017 02:10:32 -0700
Subject: [theory-seminar] University Oral Exam: Amir Abboud,
Tuesday May 30, **1:30PM**
In-Reply-To: <6baebff670b4229b6a3190fff8aeb5d9.squirrel@xenon.stanford.edu>
References: <6baebff670b4229b6a3190fff8aeb5d9.squirrel@xenon.stanford.edu>
Message-ID:
Reminder: My thesis defense is today at 1:30PM in Gates 463A.
There will be Coffee and Baklava at 1:15PM.
I hope to see you there!
Amir
On Fri, May 26, 2017 at 7:49 PM, wrote:
> University Oral Examination
>
> Title: Hardness in P
>
> Amir Abboud
> Computer Science Department, Stanford University
>
> Advised by Virginia Vassilevska Williams and Omer Reingold
>
> When: Tuesday May 30, 2017 1:30PM ***Note time change***
> Where: Gates 463A
>
> Abstract:
> The celebrated theory of NP-Hardness classifies problems into ?easy? (in
> P) and ?NP-hard? (probably not in P). However, while the easy problems do
> have polynomial time algorithms, in many cases, a polynomial runtime like
> O(n^2) or O(n^3) can be too slow. Many fundamental "easy" problems have
> resisted decades of attempts at getting truly fast algorithms, and are
> typically solved by potentially inaccurate but fast algorithms in
> practice. Such problems include Sequence Alignment, Parsing, Distance
> Computation in Graphs, and so many other problems we encounter in basic
> Algorithms courses. Can we formally explain the lack of faster algorithms
> and justify the use of heuristics, even for ?easy? problems?
>
> In this talk, I will present a theoretical framework for proving hardness
> for problems in P. Inspired by NP-Hardness, our approach is based on
> reductions which have uncovered fascinating structure among the different
> problems in P. We can now point at a small set of core problems, and say:
> For dozens of important problems in P there is no hope to get faster
> algorithms, until we know how to solve the core problems faster.
>
> _______________________________________________
> theory-seminar mailing list
> theory-seminar at lists.stanford.edu
> https://mailman.stanford.edu/mailman/listinfo/theory-seminar
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
From yuhch123 at cs.stanford.edu Tue May 30 14:47:55 2017
From: yuhch123 at cs.stanford.edu (Huacheng Yu)
Date: Tue, 30 May 2017 21:47:55 +0000
Subject: [theory-seminar] University Oral Exam: Huacheng Yu,
Wednesday May 31, 10am, Gates 104
In-Reply-To:
References:
Message-ID:
Reminder: tomorrow 10am in Gates 104. Snacks/drinks will arrive at 9:45!
Huacheng Yu ?2017?5?17??? ??3:20???
> University Oral Examination
>
> Title: Techniques for Proving Data Structure Lower Bounds
>
> Huacheng Yu
> Computer Science Department
> Stanford University
>
> Advised by Ryan Williams and Omer Reingold
>
> Wednesday, May 31, 10am, Gates 104
>
> Abstract: We present several techniques for proving the nonexistence of
> "too-
> efficient" data structures. Then we apply these techniques to prove lower
> bounds for three data structure problems: dynamic interval union, 2D
> weighted
> orthogonal range counting, and 2D parity orthogonal range counting.
>
> In this talk, I will present the lower bound results, and focus on the
> lower
> bound for the 2D parity orthogonal range counting problem. In this problem,
> the data structure maintains a set of points S in the 2D space, supporting:
> - insertion: add a point to set S;
> - query: return the parity of number of points in an axis-parallel
> rectangle.
> We show that any data structure for this problem with polylog n insertion
> time
> must have >~ log^1.5 n query time. The lower bound is proved using a new
> weak-
> simulation theorem. Roughly speaking, if the problem admits an efficient
> data
> structure, then it is possible to succinctly encode a sequence of updates
> (insertions) such that one may answer a query significantly better than
> random
> guessing based only on the encoding. The lower bound then follows from
> showing
> that the data structure problem does not have this property.
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
From tim at theory.stanford.edu Wed May 31 08:33:51 2017
From: tim at theory.stanford.edu (Tim Roughgarden)
Date: Wed, 31 May 2017 08:33:51 -0700 (PDT)
Subject: [theory-seminar] Motwani CS Theory Colloquium: Piotr Indyk (June 1)
Message-ID:
Motwani Distinguished Lectures constitute a series of theory colloquia
aimed at a broad audience. The next lecture in the series will be given
Thursday, June 1st, by Piotr Indyk from MIT. It should be a great
talk---please attend if you can!
The talk is at 4:15 PM on June 1st in the Bechtel Conference Center
(Encina Hall, 616 Serra Mall)
http://www-siepr.stanford.edu/policyforum/Encina.pdf
There will be a reception with refreshments immediately following the
talk.
Hope to see you there!
Tim
---
Title: Beyond P vs. NP: Quadratic-Time Hardness For Big Data Problems
Abstract: The theory of NP-hardness has been very successful in
identifying problems that are unlikely to be solvable in polynomial time.
However, many other important problems do have polynomial time algorithms,
but large exponents in their time bounds can make them run for days, weeks
or more. For example, quadratic time algorithms, although practical on
moderately sized inputs, can become inefficient on big data problems that
involve gigabytes or more of data. Although for many problems no
sub-quadratic time algorithms are known, any evidence of quadratic-time
hardness has remained elusive.
In this talk I will give an overview of recent research that aims to
remedy this situation. In particular, I will describe hardness results for
problems in string processing (e.g., edit distance computation or regular
expression matching) and machine learning (e.g., Support Vector Machines
or gradient computation in neural networks). All of them have polynomial
time algorithms, but despite extensive amount of research, no near-linear
time algorithms have been found. I will show that, under a natural
complexity-theoretic conjecture, such algorithms do not exist. I will also
describe how this framework has led to the development of new algorithms
for some variants of these problems.
From tim at theory.stanford.edu Wed May 31 08:33:51 2017
From: tim at theory.stanford.edu (Tim Roughgarden)
Date: Wed, 31 May 2017 08:33:51 -0700 (PDT)
Subject: [theory-seminar] [Theory Seminar] [cstheory-special] Motwani CS
Theory Colloquium: Piotr Indyk (June 1)
Message-ID:
Motwani Distinguished Lectures constitute a series of theory colloquia
aimed at a broad audience. The next lecture in the series will be given
Thursday, June 1st, by Piotr Indyk from MIT. It should be a great
talk---please attend if you can!
The talk is at 4:15 PM on June 1st in the Bechtel Conference Center
(Encina Hall, 616 Serra Mall)
http://www-siepr.stanford.edu/policyforum/Encina.pdf
There will be a reception with refreshments immediately following the
talk.
Hope to see you there!
Tim
---
Title: Beyond P vs. NP: Quadratic-Time Hardness For Big Data Problems
Abstract: The theory of NP-hardness has been very successful in
identifying problems that are unlikely to be solvable in polynomial time.
However, many other important problems do have polynomial time algorithms,
but large exponents in their time bounds can make them run for days, weeks
or more. For example, quadratic time algorithms, although practical on
moderately sized inputs, can become inefficient on big data problems that
involve gigabytes or more of data. Although for many problems no
sub-quadratic time algorithms are known, any evidence of quadratic-time
hardness has remained elusive.
In this talk I will give an overview of recent research that aims to
remedy this situation. In particular, I will describe hardness results for
problems in string processing (e.g., edit distance computation or regular
expression matching) and machine learning (e.g., Support Vector Machines
or gradient computation in neural networks). All of them have polynomial
time algorithms, but despite extensive amount of research, no near-linear
time algorithms have been found. I will show that, under a natural
complexity-theoretic conjecture, such algorithms do not exist. I will also
describe how this framework has led to the development of new algorithms
for some variants of these problems.
_______________________________________________
cstheory-special mailing list
cstheory-special at lists.stanford.edu
https://mailman.stanford.edu/mailman/listinfo/cstheory-special
_______________________________________________
Website: http://theory.stanford.edu/seminar
Visit https://mailman.stanford.edu/mailman/listinfo/algo-seminar to subscribe to or unsubscribe from this list.
From michael.kim at cs.stanford.edu Wed May 31 15:13:30 2017
From: michael.kim at cs.stanford.edu (Michael Kim)
Date: Wed, 31 May 2017 15:13:30 -0700
Subject: [theory-seminar] Theory Seminar 6/1 -- Michela Meister
Message-ID:
Hi all,
Please join us tomorrow for Theory Lunch at noon in Gates 463. Michela
Meister will be our speaker -- see her title and abstract below. Hope to
see you there!
******
Speaker: Michela Meister
Title: The Data Prism
Abstract:
We show that it is possible to leverage a small (constant-sized) dataset,
reflecting the behaviors or preferences of a cohort/demographic of
interest, to extract accurate information about this demographic from a
much larger, mixed, dataset that contains a small (and unlabelled) fraction
of datapoints from the demographic of interest. In this sense, our results
can be viewed as a ``data prism'' allowing one to extract the behavior of
specific cohorts from a large, mixed, dataset. We formally model this
problem as follows: assume an underlying set of n binary variables,
reflecting the preferences of the demographic of interest for each of n
items. Consider a large dataset in which each ``person''/datapoint
contributes estimates of the values of a subset of r of the variables: an
alpha-fraction of these datapoints (corresponding to the demographic of
interest) will produce correct (or with independent noise) estimates of the
r variables, and we make no assumptions about the remaining (1-alpha)
fraction of the datapoints---these could correspond to demographics with
similar preferences, or could correspond to random, arbitrary or even
malicious datapoints. We show that provided alpha > 1/exp(r), given
access to the preferences of the demographic of interest for a
*constant* number
of items---independent of n but dependent on alpha, r, and epsilon---we can
accurately extract the length n vector of preferences for the demographic
of interest to error epsilon, with high probability.
This result can be viewed as an instance of the semi-verified learning
model introduced by Charikar, Steinhardt, and Valiant to study robust
algorithms for extracting reliable information from large unreliable or
crowdsourced datasets. In the language of that model, we explore the
regime in which each ``evaluator'' considers as few items as possible, and
the fraction of ``reliable'' evaluators is small.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: