From hongyang at cs.stanford.edu Thu Feb 1 09:50:12 2018
From: hongyang at cs.stanford.edu (Hongyang Zhang)
Date: Thu, 01 Feb 2018 17:50:12 +0000
Subject: [theory-seminar] [theory-lunch] Irit Dinur on unique games
In-Reply-To:
References:
Message-ID:
Reminder, this is today at noon!
On Tue, Jan 30, 2018 at 10:44 AM Hongyang Zhang
wrote:
> Hi Everyone
>
> This Thursday, we are excited to have Prof. Irit Dinur, who will tell us
> about a recent progress towards proving the unique games conjecture.
> Abstract follows below.
>
> ?As before, we meet from noon to 1pm at Gates 463a.
>
> -----------------------------------------
> ?-------?
>
> Title : unique games is ?-hard
>
> I will describe recent progress towards proving the unique games
> conjecture, namely a proof that it is NP-hard to decide if a given unique
> games instance has value 1/2-o(1) or at most o(1).
> The soundness of this reduction was initially proven modulo a
> combinatorial hypothesis about the structure of non-expanding sets in the
> so-called Grassmann graph. The correctness of this hypothesis was very
> recently established by Khot Minzer and Safra, thereby completing this line
> of work.
>
> based on joint works
> ? ?
> with Khot Kindler Minzer and Safra
> ?.?
> ?------------------------------------------?
>
> ?Best,
> Hongyang?
>
>
>
> --
> Hongyang Zhang ?????
> cs.stanford.edu/people/hongyang/
>
--
--
Hongyang Zhang ?????
cs.stanford.edu/people/hongyang/
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
From hongyang at cs.stanford.edu Thu Feb 1 09:50:12 2018
From: hongyang at cs.stanford.edu (Hongyang Zhang)
Date: Thu, 01 Feb 2018 17:50:12 +0000
Subject: [theory-seminar] [theory-lunch] Irit Dinur on unique games
In-Reply-To:
References:
Message-ID:
Reminder, this is today at noon!
On Tue, Jan 30, 2018 at 10:44 AM Hongyang Zhang
wrote:
> Hi Everyone
>
> This Thursday, we are excited to have Prof. Irit Dinur, who will tell us
> about a recent progress towards proving the unique games conjecture.
> Abstract follows below.
>
> ?As before, we meet from noon to 1pm at Gates 463a.
>
> -----------------------------------------
> ?-------?
>
> Title : unique games is ?-hard
>
> I will describe recent progress towards proving the unique games
> conjecture, namely a proof that it is NP-hard to decide if a given unique
> games instance has value 1/2-o(1) or at most o(1).
> The soundness of this reduction was initially proven modulo a
> combinatorial hypothesis about the structure of non-expanding sets in the
> so-called Grassmann graph. The correctness of this hypothesis was very
> recently established by Khot Minzer and Safra, thereby completing this line
> of work.
>
> based on joint works
> ? ?
> with Khot Kindler Minzer and Safra
> ?.?
> ?------------------------------------------?
>
> ?Best,
> Hongyang?
>
>
>
> --
> Hongyang Zhang ?????
> cs.stanford.edu/people/hongyang/
>
--
--
Hongyang Zhang ?????
cs.stanford.edu/people/hongyang/
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
From silas at stanford.edu Sat Feb 3 09:16:36 2018
From: silas at stanford.edu (Shashwat Silas)
Date: Sat, 3 Feb 2018 17:16:36 +0000
Subject: [theory-seminar] Bonus Theory Seminar on Tuesday!
Message-ID:
Hi all,
We will have a bonus theory seminar on Tuesday. It will be at the usual location 463A but at 3PM. Hope you can make it!
(Tuesday!) February 6, 2018
Gates 463A, 3:00PM (Note the unusual day and time!)
Jacob Steinhardt (Stanford)
Provably Secure Machine Learning
The widespread use of machine learning systems creates a new class of computer security vulnerabilities where, rather than attacking the integrity of the software itself, malicious actors exploit the statistical nature of the learning algorithms. For instance, attackers can add fake data (e.g. by creating fake user accounts), or strategically manipulate inputs to the system once it is deployed.
So far, attempts to defend against these attacks have focused on empirical performance against known sets of attacks. I will argue that this is a fundamentally inadequate paradigm for achieving meaningful security guarantees. Instead, we need algorithms that are provably secure by design, in line with best practices for traditional computer security.
To achieve this goal, we take inspiration from robust optimization and robust statistics, but with an eye towards the security requirements of modern machine learning systems. In particular, we will develop new algorithms for robust learning in high-dimensional settings, as well as for certifiably robust optimization of non-convex models.
Thanks,
Shashwat
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
From hongyang at cs.stanford.edu Mon Feb 5 21:47:43 2018
From: hongyang at cs.stanford.edu (Hongyang Zhang)
Date: Mon, 5 Feb 2018 21:47:43 -0800
Subject: [theory-seminar] [theory-lunch] Alexandra Porter on Fractional
Repetition Codes
Message-ID:
Hi Everyone
This Thursday, Alex Porter will tell us about her recent work on
Load-Balanced Fractional Repetition Codes. Abstract follows below.
?As before, we meet from noon to 1pm, at Gates 463a.
?
--------------------------------------------
Abstract:
? ?
We introduce load-balanced fractional repetition (LBFR) codes, which are a
strengthening of fractional repetition (FR) codes. LBFR codes have the
additional property that multiple node failures can be sequentially
repaired by downloading no more than one block from any other node. This
allows for better use of the network, and can additionally reduce the
number of disk reads necessary to repair multiple nodes. We characterize
LBFR codes in terms of their adjacency graphs, and use this
characterization to present explicit constructions LBFR codes with storage
capacity comparable existing FR codes. Surprisingly, in some parameter
regimes, our constructions of LBFR codes match the parameters of the best
constructions of FR codes.
based on work with Shashwat Silas and Mary Wottters
---------------------------------------------
Best,
Hongyang?
?
--
Hongyang Zhang ?????
cs.stanford.edu/people/hongyang/
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
From hongyang at cs.stanford.edu Mon Feb 5 21:47:43 2018
From: hongyang at cs.stanford.edu (Hongyang Zhang)
Date: Mon, 5 Feb 2018 21:47:43 -0800
Subject: [theory-seminar] [theory-lunch] Alexandra Porter on Fractional
Repetition Codes
Message-ID:
Hi Everyone
This Thursday, Alex Porter will tell us about her recent work on
Load-Balanced Fractional Repetition Codes. Abstract follows below.
?As before, we meet from noon to 1pm, at Gates 463a.
?
--------------------------------------------
Abstract:
? ?
We introduce load-balanced fractional repetition (LBFR) codes, which are a
strengthening of fractional repetition (FR) codes. LBFR codes have the
additional property that multiple node failures can be sequentially
repaired by downloading no more than one block from any other node. This
allows for better use of the network, and can additionally reduce the
number of disk reads necessary to repair multiple nodes. We characterize
LBFR codes in terms of their adjacency graphs, and use this
characterization to present explicit constructions LBFR codes with storage
capacity comparable existing FR codes. Surprisingly, in some parameter
regimes, our constructions of LBFR codes match the parameters of the best
constructions of FR codes.
based on work with Shashwat Silas and Mary Wottters
---------------------------------------------
Best,
Hongyang?
?
--
Hongyang Zhang ?????
cs.stanford.edu/people/hongyang/
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
From silas at stanford.edu Tue Feb 6 10:27:38 2018
From: silas at stanford.edu (Shashwat Silas)
Date: Tue, 6 Feb 2018 18:27:38 +0000
Subject: [theory-seminar] Theory Seminar Today and Thursday
Message-ID:
Hi all,
Just a reminder that today at 3pm Jacob will give a talk for theory seminar. Also, at our usual time on Thursday 415pm at Gates 463A we will have a talk by Nima Ahmadipouranari. Both abstracts are below.
Hope to see you there!
(Tuesday!) February 6, 2018
Gates 463A, 3:00PM (Note the unusual day and time!)
Jacob Steinhardt (Stanford)
Provably Secure Machine Learning
The widespread use of machine learning systems creates a new class of computer security vulnerabilities where, rather than attacking the integrity of the software itself, malicious actors exploit the statistical nature of the learning algorithms. For instance, attackers can add fake data (e.g. by creating fake user accounts), or strategically manipulate inputs to the system once it is deployed.
So far, attempts to defend against these attacks have focused on empirical performance against known sets of attacks. I will argue that this is a fundamentally inadequate paradigm for achieving meaningful security guarantees. Instead, we need algorithms that are provably secure by design, in line with best practices for traditional computer security.
To achieve this goal, we take inspiration from robust optimization and robust statistics, but with an eye towards the security requirements of modern machine learning systems. In particular, we will develop new algorithms for robust learning in high-dimensional settings, as well as for certifiably robust optimization of non-convex models.
February 8, 2018
Gates 463A, 4:15PM
Nima Ahmadipouranari (Stanford)
Planar Graph Perfect Matching is in NC
Is matching in NC? In other words, is there a deterministic fast parallel algorithm for it? This has been an open question for over three decades, ever since the discovery of Randomized NC matching algorithms. Within this question, the case of planar graphs has remained an enigma: On the one hand, counting the number of perfect matchings is generally believed to be harder than finding one (the former is #P-complete and the latter is in P), and on the other, for planar graphs, counting has long been known to be in NC whereas finding one has resisted a solution!
The case of bipartite planar graphs was solved by Miller and Naor in 1989 via a flow-based algorithm. In 2000, Mahajan and Varadarajan gave an elegant way of using counting matchings to finding one, hence giving a different NC algorithm.
However, non-bipartite planar graphs still didn't yield: the stumbling block being tight odd cuts. Interestingly enough, these are also a key to the solution: a balanced odd tight cut leads to a straight-forward divide and conquer NC algorithm. The remaining task is to find such a cut in NC. This requires several algorithmic ideas, such as finding a point in the interior of the minimum weight face of the perfect matching polytope, and uncrossing tight odd cuts.
Paper available at: https://arxiv.org/pdf/1709.07822.pdf
Joint work with Vijay Vazirani.
Thanks,
Shashwat
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
From clc2200 at columbia.edu Wed Feb 7 08:32:22 2018
From: clc2200 at columbia.edu (=?UTF-8?Q?Cl=c3=a9ment_Canonne?=)
Date: Wed, 7 Feb 2018 08:32:22 -0800
Subject: [theory-seminar] TCS+ talk: Wednesday, February 14, Dor Minzer,
Tel-Aviv University
Message-ID: <48ae92b4-e7cc-c9b8-81f8-6d02b419ee7d@cs.columbia.edu>
Hi all,
The second TCS+ talk of the semester is happening next Wednesday, with
the perfect way to spend Valentine's day: learning about the recent
breakthrough on the 2-to-2 conjecture by Khot, Minzer, and Safra, and
this overall line of work. (Also, a great continuation to Irit's talk on
this topic from last week!)
Unfortunately, I won't be here myself to set up the talk -- if one of
you is interested in doing so (it's very simple) so that we can have a
viewing in the department, shoot me an email!
(Title and abstract below)
Best,
-- Cl?ment
-------------------------------
Speaker: Dor Minzer (Tel-Aviv University)
Title: 2-to-2 Games via expansion on the Grassmann Graph
Abstract: A fundamental goal in the theory of PCPs asks whether a
special type of PCP, namely 2-to-2 Games, exists. This is a variant of
Khot's well-known Unique Games conjecture.
In this talk we will discuss a recent line of study establishing the
2-to-2 games conjecture, albeit with imperfect completeness.
At the heart of the approach lies an object called the Grassmann Graph,
and a certain linearity test on it.
This leads to the study of edge expansion in this graph, and in
particular, the study of (small) sets of vertices in the Grassmann
Graph, whose edge expansion is bounded away from 1.
Based on joint works with Irit Dinur, Subhash Khot, Guy Kindler and Muli
Safra
From gvaliant at cs.stanford.edu Thu Feb 8 09:28:17 2018
From: gvaliant at cs.stanford.edu (Gregory Valiant)
Date: Thu, 8 Feb 2018 09:28:17 -0800
Subject: [theory-seminar] post icml-deadline happy hour!
Message-ID:
Hi Friends,
Due to the fortuitous timing of the ICML deadline (4pm tomorrow---I
think?), we'll have a post-deadline happy hour in Gates 463a starting
around 4:15pm. I'll bring some beer, and maybe some snacks. (And of
course, submitting a paper is NOT a requisite for attending the happy hour.)
Hope to see you tomorrow,
-g
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
From ccanonne at cs.stanford.edu Thu Feb 8 10:04:36 2018
From: ccanonne at cs.stanford.edu (=?UTF-8?Q?Cl=c3=a9ment_Canonne?=)
Date: Thu, 8 Feb 2018 10:04:36 -0800
Subject: [theory-seminar] post icml-deadline happy hour!
In-Reply-To:
References:
Message-ID: <28fd8b85-02dc-9985-5aa8-51373b86239d@cs.stanford.edu>
Hi all,
Also, I know it's late, but for what it's worth: if one or two of you
are submitting something and are looking for some feedback on the
intro/first pages, I can try and play the role of the not-so-informed
reviewer.
Best,
Cl?ment
On 02/08/2018 09:28 AM, Gregory Valiant wrote:
> Hi Friends,
> Due to the fortuitous timing of the ICML deadline (4pm tomorrow---I
> think?), we'll have a post-deadline happy hour in Gates 463a starting
> around 4:15pm.? ?I'll bring some beer, and maybe some snacks.? ?(And of
> course, submitting a paper is NOT a requisite for attending the happy hour.)
>
> Hope to see you tomorrow,
> -g
>
>
>
>
>
> _______________________________________________
> theory-seminar mailing list
> theory-seminar at lists.stanford.edu
> https://mailman.stanford.edu/mailman/listinfo/theory-seminar
>
From gvaliant at cs.stanford.edu Thu Feb 8 10:43:30 2018
From: gvaliant at cs.stanford.edu (Gregory Valiant)
Date: Thu, 8 Feb 2018 10:43:30 -0800
Subject: [theory-seminar] post icml-deadline happy hour!
In-Reply-To:
References:
Message-ID:
Slight change of plan---if its warm outside, we'll have the post-deadline
party on the patio out back, otherwise we will be in Gates 463a.
Cheers,
-g
On Thu, Feb 8, 2018 at 9:28 AM, Gregory Valiant
wrote:
> Hi Friends,
> Due to the fortuitous timing of the ICML deadline (4pm tomorrow---I
> think?), we'll have a post-deadline happy hour in Gates 463a starting
> around 4:15pm. I'll bring some beer, and maybe some snacks. (And of
> course, submitting a paper is NOT a requisite for attending the happy hour.)
>
> Hope to see you tomorrow,
> -g
>
>
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
From silas at stanford.edu Thu Feb 8 14:32:14 2018
From: silas at stanford.edu (Shashwat Silas)
Date: Thu, 8 Feb 2018 22:32:14 +0000
Subject: [theory-seminar] Theory Seminar Today
Message-ID:
Hi all,
There will be theory seminar today at 415pm in Gates 463A. See you there!
February 8, 2018
Gates 463A, 4:15PM
Nima Ahmadipouranari (Stanford)
Planar Graph Perfect Matching is in NC
Is matching in NC? In other words, is there a deterministic fast parallel algorithm for it? This has been an open question for over three decades, ever since the discovery of Randomized NC matching algorithms. Within this question, the case of planar graphs has remained an enigma: On the one hand, counting the number of perfect matchings is generally believed to be harder than finding one (the former is #P-complete and the latter is in P), and on the other, for planar graphs, counting has long been known to be in NC whereas finding one has resisted a solution!
The case of bipartite planar graphs was solved by Miller and Naor in 1989 via a flow-based algorithm. In 2000, Mahajan and Varadarajan gave an elegant way of using counting matchings to finding one, hence giving a different NC algorithm.
However, non-bipartite planar graphs still didn't yield: the stumbling block being tight odd cuts. Interestingly enough, these are also a key to the solution: a balanced odd tight cut leads to a straight-forward divide and conquer NC algorithm. The remaining task is to find such a cut in NC. This requires several algorithmic ideas, such as finding a point in the interior of the minimum weight face of the perfect matching polytope, and uncrossing tight odd cuts.
Paper available at: https://arxiv.org/pdf/1709.07822.pdf
Joint work with Vijay Vazirani.
Shashwat
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
From rayyli at stanford.edu Tue Feb 13 13:16:59 2018
From: rayyli at stanford.edu (Ray Li)
Date: Tue, 13 Feb 2018 21:16:59 +0000
Subject: [theory-seminar] TCS+ talk: Wednesday, February 14, Dor Minzer,
Tel-Aviv University
In-Reply-To: <48ae92b4-e7cc-c9b8-81f8-6d02b419ee7d@cs.columbia.edu>
References: <48ae92b4-e7cc-c9b8-81f8-6d02b419ee7d@cs.columbia.edu>
Message-ID:
Hi All,
We will have a viewing of the TCS+ talk tomorrow at 10am in Gates 463A. I've recopied the title and abstract below.
Unfortunately, all the google hangouts spots were taken, so we will watch it non-interactively.
Best,
Ray
-------------------------------
Speaker: Dor Minzer (Tel-Aviv University)
Title: 2-to-2 Games via expansion on the Grassmann Graph
Abstract: A fundamental goal in the theory of PCPs asks whether a
special type of PCP, namely 2-to-2 Games, exists. This is a variant of
Khot's well-known Unique Games conjecture.
In this talk we will discuss a recent line of study establishing the
2-to-2 games conjecture, albeit with imperfect completeness.
At the heart of the approach lies an object called the Grassmann Graph,
and a certain linearity test on it.
This leads to the study of edge expansion in this graph, and in
particular, the study of (small) sets of vertices in the Grassmann
Graph, whose edge expansion is bounded away from 1.
Based on joint works with Irit Dinur, Subhash Khot, Guy Kindler and Muli
Safra
________________________________
From: theory-seminar on behalf of Cl?ment Canonne
Sent: Wednesday, February 7, 2018 8:32:22 AM
To: theory-seminar at mailman.stanford.edu
Subject: [theory-seminar] TCS+ talk: Wednesday, February 14, Dor Minzer, Tel-Aviv University
Hi all,
The second TCS+ talk of the semester is happening next Wednesday, with
the perfect way to spend Valentine's day: learning about the recent
breakthrough on the 2-to-2 conjecture by Khot, Minzer, and Safra, and
this overall line of work. (Also, a great continuation to Irit's talk on
this topic from last week!)
Unfortunately, I won't be here myself to set up the talk -- if one of
you is interested in doing so (it's very simple) so that we can have a
viewing in the department, shoot me an email!
(Title and abstract below)
Best,
-- Cl?ment
-------------------------------
Speaker: Dor Minzer (Tel-Aviv University)
Title: 2-to-2 Games via expansion on the Grassmann Graph
Abstract: A fundamental goal in the theory of PCPs asks whether a
special type of PCP, namely 2-to-2 Games, exists. This is a variant of
Khot's well-known Unique Games conjecture.
In this talk we will discuss a recent line of study establishing the
2-to-2 games conjecture, albeit with imperfect completeness.
At the heart of the approach lies an object called the Grassmann Graph,
and a certain linearity test on it.
This leads to the study of edge expansion in this graph, and in
particular, the study of (small) sets of vertices in the Grassmann
Graph, whose edge expansion is bounded away from 1.
Based on joint works with Irit Dinur, Subhash Khot, Guy Kindler and Muli
Safra
_______________________________________________
theory-seminar mailing list
theory-seminar at lists.stanford.edu
https://mailman.stanford.edu/mailman/listinfo/theory-seminar
theory-seminar Info Page - Stanford University
mailman.stanford.edu
For scheduling meetings/lunches/lectures/talks for the CS Theory group. To see the collection of prior postings to the list, visit the theory-seminar Archives.
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
From silas at stanford.edu Tue Feb 13 16:35:03 2018
From: silas at stanford.edu (Shashwat Silas)
Date: Wed, 14 Feb 2018 00:35:03 +0000
Subject: [theory-seminar] Theory Seminar on Thursday
Message-ID:
Hi all,
Theory seminar will be on Thursday at 415pm in Gates 463A. If you'd like to meet with the speaker please email marykw at stanford.edu.
See you there!
February 15, 2018
Gates 463A, 4:15PM
Christian Borgs (MSR)
Graphons: From Graph Limits to Non-Parametric Estimation and Recommendation Systems
Graphons were invented to model the limit of large, dense graphs. While this led to interesting applications in combinatorics and property testing, most applications require limits of sparse graphs. In this talk, I will review recent progress on graph limits for sparse graphs, and then discuss a couple of applications: non-parametric modelling and estimation of sparse graphs, and recommendation systems where the matrix of known ratings is so sparse that two typical users have never rated the same item, making standard similarity based recommendation algorithms challenging. This is joint work with Jennifer Chayes, Henry Cohn, and many others.
Shashwat
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
From hongyang at cs.stanford.edu Wed Feb 14 08:54:57 2018
From: hongyang at cs.stanford.edu (Hongyang Zhang)
Date: Wed, 14 Feb 2018 08:54:57 -0800
Subject: [theory-seminar] [theory-lunch] Nima Anari on Matroids,
Entropy and Log-concavity
Message-ID:
Hi Everyone
This Thursday, Nima Anari will tell us about his work on *Matroids, Entropy
and Log-concavity*. See abstract below.
As before, we meet from noon to 1pm, at Gates 463a.
------------------------------
Abstract:
Computational connections between optimization and counting appear in many
places in computer science. In this talk I will introduce a general
framework for approximate counting in discrete problems, based on convex
optimization. The connection goes through subadditivity of the entropy. I
will show that surprisingly for a large class of discrete structures,
Matroids and their intersections, one can prove an ?approximate
superadditivity? of the entropy, resulting in nearly tight deterministic
polynomial time approximation algorithms. For Matroids we obtain a
square-root approximation, and for intersections of two Matroids we obtain
2^{O(rank)}-approximations.
The key to showing ?approximate superadditivity? is a new structural
characterization of Matroids: The basis generating polynomial of a Matroid
is log concave as a function over the positive orthant. The proof relies on
?combinatorial Hodge theory? results proved by Huh et al. and Adiprasito et
al. We also show that the support of any homogeneous multilinear log
concave polynomial must be a Matroid, thus getting a characterization of
Matroids in terms of log concave polynomials.
Time permitting I will mention another consequence of log concavity for
Matroids: approximate negative correlation in a uniformly sampled basis.
based on joint work with Shayan Oveis Gharan and Cynthia Vinzant.
--------------------------------------
Best,
Hongyang
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
From hongyang at cs.stanford.edu Wed Feb 14 08:54:57 2018
From: hongyang at cs.stanford.edu (Hongyang Zhang)
Date: Wed, 14 Feb 2018 08:54:57 -0800
Subject: [theory-seminar] [theory-lunch] Nima Anari on Matroids,
Entropy and Log-concavity
Message-ID:
Hi Everyone
This Thursday, Nima Anari will tell us about his work on *Matroids, Entropy
and Log-concavity*. See abstract below.
As before, we meet from noon to 1pm, at Gates 463a.
------------------------------
Abstract:
Computational connections between optimization and counting appear in many
places in computer science. In this talk I will introduce a general
framework for approximate counting in discrete problems, based on convex
optimization. The connection goes through subadditivity of the entropy. I
will show that surprisingly for a large class of discrete structures,
Matroids and their intersections, one can prove an ?approximate
superadditivity? of the entropy, resulting in nearly tight deterministic
polynomial time approximation algorithms. For Matroids we obtain a
square-root approximation, and for intersections of two Matroids we obtain
2^{O(rank)}-approximations.
The key to showing ?approximate superadditivity? is a new structural
characterization of Matroids: The basis generating polynomial of a Matroid
is log concave as a function over the positive orthant. The proof relies on
?combinatorial Hodge theory? results proved by Huh et al. and Adiprasito et
al. We also show that the support of any homogeneous multilinear log
concave polynomial must be a Matroid, thus getting a characterization of
Matroids in terms of log concave polynomials.
Time permitting I will mention another consequence of log concavity for
Matroids: approximate negative correlation in a uniformly sampled basis.
based on joint work with Shayan Oveis Gharan and Cynthia Vinzant.
--------------------------------------
Best,
Hongyang
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
From reingold at stanford.edu Thu Feb 15 08:21:38 2018
From: reingold at stanford.edu (Omer Reingold)
Date: Thu, 15 Feb 2018 08:21:38 -0800
Subject: [theory-seminar] Theory Dish blog
Message-ID:
Hi Everyone,
For the sake of 1st year students and anybody new to the list, I want to
remind you that our blog https://theorydish.blog/ belongs to all of us and
you are all welcomed to contribute.
In particular, the sharing tips project https://theorydish.blog/tag/tips/ is
open to contributions from all.
Best wishes,
Omer
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
From moses at cs.stanford.edu Thu Feb 15 10:11:40 2018
From: moses at cs.stanford.edu (Moses Charikar)
Date: Thu, 15 Feb 2018 10:11:40 -0800
Subject: [theory-seminar] Fwd: summer internships at Google
In-Reply-To:
References:
Message-ID:
Theory folks,
The Market Algorithms research team at Google is looking for summer interns
and is also interested in hiring for full-time positions. See the message
below from Gagan Aggarwal, theory group alum and manager of the group.
--Moses
PS. Apologies for the wide distribution. As far as I can tell, we don't
have a Stanford-only theory list.
---------- Forwarded message ----------
From: Gagan Aggarwal
The Market Algorithms research team in Mountain View is looking for interns
for the coming summer. As you know, our team works on topics at the
intersection of Algorithms and Markets, including Auction and Mechanism
design and Online Algorithms. If you have any students who might be
interested, please have them send their resume to me (gagana at google.com).
Our team is also hiring for full-time positions. If you know any graduating
students who might be interested, do let me know.
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
From silas at stanford.edu Thu Feb 15 15:32:47 2018
From: silas at stanford.edu (Shashwat Silas)
Date: Thu, 15 Feb 2018 23:32:47 +0000
Subject: [theory-seminar] Theory Seminar in 45 mins
Message-ID:
Hi all,
Just a reminder that theory seminar will be in 45 minutes in Gates 463A.
Hope you can make it,
Shashwat
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
From silas at stanford.edu Mon Feb 19 14:07:02 2018
From: silas at stanford.edu (Shashwat Silas)
Date: Mon, 19 Feb 2018 22:07:02 +0000
Subject: [theory-seminar] Theory Seminar on Thursday
Message-ID:
Hi all,
Theory seminar will be on Thursday at 415pm in Gates 463A. If you'd like to meet with the speaker please email marykw at stanford.edu. Hope you can make it!
Tom Gur (Berkeley)
Relaxed Locally Correctable Codes
Locally decodable codes (LDCs) and locally correctable codes (LCCs) are error-correcting codes in which individual bits of the message and codeword, respectively, can be recovered by querying only few bits from a noisy codeword. These codes have found numerous applications both in theory and in practice.
A natural relaxation of LDCs, introduced by Ben-Sasson et al. (SICOMP, 2006), allows the decoder to reject (i.e., refuse to answer) in case it detects that the codeword is corrupt. They call such a decoder a relaxed decoder and construct a constant-query relaxed LDC with almost-linear blocklength, which is sub-exponentially better than what is known for (full-fledged) LDCs in the constant-query regime.
We consider an analogous relaxation for local correction. Thus, a relaxed local corrector reads only few bits from a (possibly) corrupt codeword and either recovers the desired bit of the codeword, or rejects in case it detects a corruption.
We give two constructions of relaxed LCCs in two regimes, where the first optimizes the query complexity and the second optimizes the rate:
1. Constant Query Complexity: A relaxed LCC with polynomial blocklength whose corrector only reads a constant number of bits of the codeword. This is a sub-exponential improvement over the best constant query (full-fledged) LCCs that are known.
2. Constant Rate: A relaxed LCC with constant rate (i.e., linear blocklength) with quasi-polylogarithmic query complexity. This is a nearly sub-exponential improvement over the query complexity of a recent (full-fledged) constant-rate LCC of Kopparty et al. (STOC, 2016).
Joint work with Govind Ramnarayan and Ron Rothblum.
Thanks,
Shashwat
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
From hongyang at cs.stanford.edu Wed Feb 21 00:35:00 2018
From: hongyang at cs.stanford.edu (Hongyang Zhang)
Date: Wed, 21 Feb 2018 00:35:00 -0800
Subject: [theory-seminar] [theory-lunch] Hongyang Zhang on Implicit
regularization
Message-ID:
Hi Everyone
This Thursday, I will present recent work on *Algorithmic Regularization in
Over-parameterized Models. *See abstract below.
As before, we meet from noon to 1:00pm, at Gates 463a.
-----------------------------------------------
We show that gradient descent provides an implicit regularization effect in
the learning of over-parameterized matrix factorization models.
Concretely, we show that given
?~ ?
dr^2 random linear measurements of a rank r positive semidefinite matrix
?X^*?
, we can recover
X^*
by parameterizing it by UU^
?T?
with U
? being? d by
d and minimizing the squared loss, even if r
? is much less than?
d. We prove that starting from a small initialization, gradient descent
recovers X
?^*?
in \sqrt{r} iterations approximately.
?Joint work with Yuanzhi Li and Tengyu Ma?
---------------------------------------------------
Best,
Hongyang
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
From hongyang at cs.stanford.edu Wed Feb 21 00:35:00 2018
From: hongyang at cs.stanford.edu (Hongyang Zhang)
Date: Wed, 21 Feb 2018 00:35:00 -0800
Subject: [theory-seminar] [theory-lunch] Hongyang Zhang on Implicit
regularization
Message-ID:
Hi Everyone
This Thursday, I will present recent work on *Algorithmic Regularization in
Over-parameterized Models. *See abstract below.
As before, we meet from noon to 1:00pm, at Gates 463a.
-----------------------------------------------
We show that gradient descent provides an implicit regularization effect in
the learning of over-parameterized matrix factorization models.
Concretely, we show that given
?~ ?
dr^2 random linear measurements of a rank r positive semidefinite matrix
?X^*?
, we can recover
X^*
by parameterizing it by UU^
?T?
with U
? being? d by
d and minimizing the squared loss, even if r
? is much less than?
d. We prove that starting from a small initialization, gradient descent
recovers X
?^*?
in \sqrt{r} iterations approximately.
?Joint work with Yuanzhi Li and Tengyu Ma?
---------------------------------------------------
Best,
Hongyang
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
From ccanonne at cs.stanford.edu Wed Feb 21 07:31:35 2018
From: ccanonne at cs.stanford.edu (=?UTF-8?Q?Cl=c3=a9ment_Canonne?=)
Date: Wed, 21 Feb 2018 07:31:35 -0800
Subject: [theory-seminar] TCS+ talk: Wednesday, February 28, Sanjam Garg,
UC Berkeley
In-Reply-To: <485ea247-b367-2bdb-f9b1-b42b3bde1278@cms.caltech.edu>
References: <485ea247-b367-2bdb-f9b1-b42b3bde1278@cms.caltech.edu>
Message-ID: <32a82926-2552-4655-f566-b66ed7c639eb@cs.stanford.edu>
Hi everyone,
Slightly departing from my original announcement (Michael Kearns' talk
is postponed until later in the semester), the next TCS+ talk will take
place next Wednesday (28th) at *10am*, and be about a recent work of
Sanjam Garg and Nico D?ttling on "Identity-Based Encryption from the
Diffie-Hellman Assumption" (abstract below).
Sanjam will be presenting online, while we have breakfast (starting at
9:55) and can ask questions from the comfort of Gates 463A.
Hope to see you on Wednesday,
-- Cl?ment
-------------------------------
Speaker: Sanjam Garg (UC Berkeley)
Title: Identity-Based Encryption from the Diffie-Hellman Assumption
Abstract: In this talk, I will describe new constructions of
identity-based encryption based on the hardness of the Diffie-Hellman
(without using groups with pairings) Problem. Previously, constructions
based on this assumption were believed to be impossible. Our
construction is based on new techniques that bypass the known
impossibility results using garbled circuits that make a non-black-box
use of the underlying cryptographic primitives.
(Based on joint work with Nico D?ttling.)
From silas at stanford.edu Thu Feb 22 15:50:54 2018
From: silas at stanford.edu (Shashwat Silas)
Date: Thu, 22 Feb 2018 23:50:54 +0000
Subject: [theory-seminar] Theory Seminar in 25 minutes!
Message-ID:
Hi all,
Theory seminar will be in Gates 463A in 25 minutes. Hope to see you there!
Thanks,
Shashwat
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
From hongyang at cs.stanford.edu Mon Feb 26 23:31:39 2018
From: hongyang at cs.stanford.edu (Hongyang Zhang)
Date: Mon, 26 Feb 2018 23:31:39 -0800
Subject: [theory-seminar] [theory-lunch] Paris Syminelakis on hashing
Message-ID:
Hi Everyone
This Thursday, Paris Syminelakis will tell us about
?his recent work on ?
*Integrating Functions via Distance Sensitive Hashing*. See abstract below
?.?
?As before, we meet from noon to 1pm, at Gates 463a.?
--------------------------------------------------------
?Abstract:
?
Many popular computational primitives in Machine learning and Optimization
involve summing up a pairwise function over a large data-set. Such
primitives include computing the Kernel Density, Partition Function, and
Gradients of Empirical Loss functions. These procedures require computing
a discrete integral that depends on vector $y$ (query), that is unknown a
priori or might change with time. We describe a technique that utilizes
``Distance Sensitive Hashing" to provide fast approximations to such
integrals that in certain cases results in $\sqrt{n}$ speedups over the
previously best known data structures. Our technique extends some recent
results from FOCS'17 and essentially amounts to constructing for a certain
class of functions an efficient Monte Carlo-version of the Riemann Integral
over the Unit Sphere.
?B?
ased on joint work with Moses Charikar.
---------------------------------------------------
?Best,
Hongyang?
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
From hongyang at cs.stanford.edu Mon Feb 26 23:31:39 2018
From: hongyang at cs.stanford.edu (Hongyang Zhang)
Date: Mon, 26 Feb 2018 23:31:39 -0800
Subject: [theory-seminar] [theory-lunch] Paris Syminelakis on hashing
Message-ID:
Hi Everyone
This Thursday, Paris Syminelakis will tell us about
?his recent work on ?
*Integrating Functions via Distance Sensitive Hashing*. See abstract below
?.?
?As before, we meet from noon to 1pm, at Gates 463a.?
--------------------------------------------------------
?Abstract:
?
Many popular computational primitives in Machine learning and Optimization
involve summing up a pairwise function over a large data-set. Such
primitives include computing the Kernel Density, Partition Function, and
Gradients of Empirical Loss functions. These procedures require computing
a discrete integral that depends on vector $y$ (query), that is unknown a
priori or might change with time. We describe a technique that utilizes
``Distance Sensitive Hashing" to provide fast approximations to such
integrals that in certain cases results in $\sqrt{n}$ speedups over the
previously best known data structures. Our technique extends some recent
results from FOCS'17 and essentially amounts to constructing for a certain
class of functions an efficient Monte Carlo-version of the Riemann Integral
over the Unit Sphere.
?B?
ased on joint work with Moses Charikar.
---------------------------------------------------
?Best,
Hongyang?
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
From ccanonne at cs.stanford.edu Tue Feb 27 10:06:35 2018
From: ccanonne at cs.stanford.edu (=?UTF-8?Q?Cl=c3=a9ment_Canonne?=)
Date: Tue, 27 Feb 2018 10:06:35 -0800
Subject: [theory-seminar] TCS+ talk: Wednesday, February 28, Sanjam Garg,
UC Berkeley
In-Reply-To: <32a82926-2552-4655-f566-b66ed7c639eb@cs.stanford.edu>
References: <485ea247-b367-2bdb-f9b1-b42b3bde1278@cms.caltech.edu>
<32a82926-2552-4655-f566-b66ed7c639eb@cs.stanford.edu>
Message-ID:
Reminder: this is *tomorrow*, (10-eps)am.
Best,
-- Cl?ment
On 02/21/2018 07:31 AM, Cl?ment Canonne wrote:
> Hi everyone,
>
> Slightly departing from my original announcement (Michael Kearns' talk
> is postponed until later in the semester), the next TCS+ talk will take
> place next Wednesday (28th) at *10am*, and be about a recent work of
> Sanjam Garg and Nico D?ttling on "Identity-Based Encryption from the
> Diffie-Hellman Assumption" (abstract below).
>
> Sanjam will be presenting online, while we have breakfast (starting at
> 9:55) and can ask questions from the comfort of Gates 463A.
>
> Hope to see you on Wednesday,
>
> -- Cl?ment
>
>
> -------------------------------
> Speaker: Sanjam Garg (UC Berkeley)
> Title: Identity-Based Encryption from the Diffie-Hellman Assumption
>
> Abstract: In this talk, I will describe new constructions of
> identity-based encryption based on the hardness of the Diffie-Hellman
> (without using groups with pairings) Problem. Previously, constructions
> based on this assumption were believed to be impossible. Our
> construction is based on new techniques that bypass the known
> impossibility results using garbled circuits that make a non-black-box
> use of the underlying cryptographic primitives.
>
> (Based on joint work with Nico D?ttling.)
>
> _______________________________________________
> theory-seminar mailing list
> theory-seminar at lists.stanford.edu
> https://mailman.stanford.edu/mailman/listinfo/theory-seminar
From ccanonne at cs.stanford.edu Wed Feb 28 15:03:28 2018
From: ccanonne at cs.stanford.edu (=?UTF-8?Q?Cl=c3=a9ment_Canonne?=)
Date: Wed, 28 Feb 2018 15:03:28 -0800
Subject: [theory-seminar] Theory Happy Hour: S01E02 (Arrival)
Message-ID: <5c1d3bc2-c520-fe6b-5c6a-f0c2ed27e63d@cs.stanford.edu>
Hi everyone,
Now that the Winter Olympics are over, the regular programming can
resume! (Also, if someone has a clue about the rules of curling, let me
know.)
*tl;dr:* March 07th, 6pm, @ Gates 463
-----------------------
Next Wednesday, seize the chance to have a beer or drink of your choice*
(and snacks/food**) with your fellow grad students, faculty, postdocs,
*and two of the incoming faculty* in the middle of the department on
Wednesday, January 24th***.
Namely, this event will be attended (among other noteworthy attendees)
by both Li-Yang Tan and Tengyu Ma, two of the new professors joining in
Fall. Hence the title of this email.
Event sponsored by the faculty (Mary and Greg, within reasonable bounds).
* Based on availability.
** Please, email me any suggestion or constraint by Sunday, so that I
can accommodate diet restrictions unless they are just clearly to drive
me crazy (paleo-vegan?)
*** If the weather is nice enough, we can even think of moving outside.
The world is our salt-water bivalve mollusc.
Best,
--
Cl?ment
-------------- next part --------------
A non-text attachment was scrubbed...
Name: arrival.jpg
Type: image/jpeg
Size: 47101 bytes
Desc: not available
URL: