From ofirgeri at stanford.edu Mon May 6 15:37:46 2019
From: ofirgeri at stanford.edu (Ofir Geri)
Date: Mon, 6 May 2019 22:37:46 +0000
Subject: [theory-seminar] Theory Seminar (5/10): Kira Goldner
Message-ID:
Hi all,
This week in the theory seminar, Kira Goldner (University of Washington) will give a talk on Mechanism Design under Interdependent Valuations (see abstract below). The talk will be as usual on Friday, 3:00pm in Gates 463A.
The abstracts of past and upcoming seminar talks is also available on the theory seminar webpage:
http://theory.stanford.edu/seminar/
Hope to see you there!
Ofir
Mechanism Design under Interdependent Valuations
Speaker: Kira Goldner (University of Washington)
We study buyers with interdependent valuations: where one buyer's private information about an item impacts how much another buyer is willing to pay for it. In this setting, if a buyer misreports his own private information, he can impact not only what the auctioneer believes his own value is, but also what the auctioneer believes regarding other buyers' values as well, allowing him to essentially corrupt their values. As a result, the usual mechanism design tricks fall short, and welfare maximization is notoriously difficult. Almost all prior work in this setting requires a very strong "single-crossing" condition on the valuation functions in order to obtain any positive results.
We introduce two more natural notions -- first, a relaxed, parameterized notion of single-crossing, and second, a completely unrelated notion, one of submodularity over the private information of buyers -- that each separately enable good approximation guarantees to optimal welfare. These conditions, combined with the lens of approximation, allow us to go beyond not only the restrictive single-crossing notion, but to go far beyond the single-item setting and to give positive results for even the very general setting of combinatorial auctions.
Joint work with Alon Eden, Michal Feldman, Amos Fiat, and Anna Karlin.
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
From wyma at stanford.edu Mon May 6 16:53:59 2019
From: wyma at stanford.edu (Weiyun Ma)
Date: Mon, 6 May 2019 23:53:59 +0000
Subject: [theory-seminar] =?iso-8859-1?q?Theory_Lunch_5/9_--_Cl=E9ment_Can?=
=?iso-8859-1?q?onne?=
Message-ID:
Hi all,
This Thursday at theory lunch, Cl?ment will tell us about "Distributed goodness-of-fit: when you can't talk much and have little in common." (See abstract below.)
As usual, please join us from noon to 1pm at 463A.
----------------------------------------------------------
Distributed goodness-of-fit: when you can't talk much and have little in common
Speaker: Cl?ment Canonne
Consider n low-battery sensors, spread across the city, performing each an independent measurement in order to allow a central server to detect changes in the distribution of temperature. For concreteness, suppose the measurements are quantized and take values in {1,2,..,k}: due to the battery and throughput constraints, the sensors can only send L << log k bits each, in one-way fashion, to the server.
Previous work of Acharya, Canonne, and Tyagi [ACT'18] established optimal protocols to perform distributed goodness-of-fit (identity testing) task under these communication constraints; further, they showed a quantitative separation between private-coin protocols (where each sensors can only use its own private coin flips) and public-coin ones (where they share a common random seed, e.g., hardcoded or broadcast by the central server ahead of time). In this work, we refine the question and ask how this quantitative gap behaves with the amount
of shared random bits, which we see as an additional resource: namely, we fully characterize the optimal sample complexity of distributed identity testing as a function of k, L, and the number of shared random bits s.
A key aspect of our protocols is a derandomization lemma which we strongly believe to be of independent interest, as well as a generalization of the lower bound framework of [ACT'18] to account for limited shared randomness.
Joint work with Jayadev Acharya, Yanjun Han, Ziteng Sun, and Himanshyu Tyagi.
----------------------------------------------------------
Best,
Anna
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
From wyma at stanford.edu Mon May 6 16:53:59 2019
From: wyma at stanford.edu (Weiyun Ma)
Date: Mon, 6 May 2019 23:53:59 +0000
Subject: [theory-seminar] =?iso-8859-1?q?Theory_Lunch_5/9_--_Cl=E9ment_Can?=
=?iso-8859-1?q?onne?=
Message-ID:
Hi all,
This Thursday at theory lunch, Cl?ment will tell us about "Distributed goodness-of-fit: when you can't talk much and have little in common." (See abstract below.)
As usual, please join us from noon to 1pm at 463A.
----------------------------------------------------------
Distributed goodness-of-fit: when you can't talk much and have little in common
Speaker: Cl?ment Canonne
Consider n low-battery sensors, spread across the city, performing each an independent measurement in order to allow a central server to detect changes in the distribution of temperature. For concreteness, suppose the measurements are quantized and take values in {1,2,..,k}: due to the battery and throughput constraints, the sensors can only send L << log k bits each, in one-way fashion, to the server.
Previous work of Acharya, Canonne, and Tyagi [ACT'18] established optimal protocols to perform distributed goodness-of-fit (identity testing) task under these communication constraints; further, they showed a quantitative separation between private-coin protocols (where each sensors can only use its own private coin flips) and public-coin ones (where they share a common random seed, e.g., hardcoded or broadcast by the central server ahead of time). In this work, we refine the question and ask how this quantitative gap behaves with the amount
of shared random bits, which we see as an additional resource: namely, we fully characterize the optimal sample complexity of distributed identity testing as a function of k, L, and the number of shared random bits s.
A key aspect of our protocols is a derandomization lemma which we strongly believe to be of independent interest, as well as a generalization of the lower bound framework of [ACT'18] to account for limited shared randomness.
Joint work with Jayadev Acharya, Yanjun Han, Ziteng Sun, and Himanshyu Tyagi.
----------------------------------------------------------
Best,
Anna
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
From shivamgarg at stanford.edu Tue May 7 20:37:59 2019
From: shivamgarg at stanford.edu (Shivam Garg)
Date: Wed, 8 May 2019 03:37:59 +0000
Subject: [theory-seminar] Algorithms and Friends Lunch,
Monday (May 13) at Gates 463A
In-Reply-To:
References:
Message-ID:
Hi everyone,
Avidit Acharya will be giving a talk on coming Monday (May 13th), at noon, in Gates 463A.
Title: Matching Refugees to Host Country Locations Based on Preferences and Outcomes (with Kirk Bansak and Jens Hainmueller)
Abstract: Facilitating the integration of refugees has become a major policy challenge in many host countries in the context of the global displacement crisis. One of the first policy decisions host countries make in the resettlement process is the assignment of refugees to locations within the country. We develop a mechanism to match refugees to locations in a way that takes into account their expected integration outcomes and their preferences over where to be settled. Our proposal is based on a priority mechanism that allows the government first to specify a threshold g for the minimum level of expected integration success that should be achieved. Refugees are then matched to locations based on their preferences subject to meeting the government's specified threshold. The mechanism is both strategy-proof and constrained efficient in that it always generates a matching that is not Pareto dominated by any other matching that respects the government's threshold. We demonstrate our approach using simulations and a real-world application to refugee data from the United States.
Speaker Bio: Avidit (Avi) Acharya is an assistant professor of political science at Stanford. He is a formal political theorist and political economist whose work ranges across a diverse set of topics including voting theory, bargaining theory, principal-agent theory, behavioral political economy, distributive politics, and long run development. Avi?s papers have been published (or are forthcoming) in the leading journals of political science, including the American Political Science Review, American Journal of Political Science, and Journal of Politics, as well as the top journals in economic theory, including Econometrica, Journal of Economic Theory, and Games and Economic Behavior. Avi earned his PhD in political economy from Princeton University, and taught in the political science and economics departments of the University of Rochester before coming to Stanford.
Thanks,
Shivam
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
-------------- next part --------------
An embedded and charset-unspecified text was scrubbed...
Name: ATT00001.txt
URL:
From moses at cs.stanford.edu Tue May 7 23:08:02 2019
From: moses at cs.stanford.edu (Moses Charikar)
Date: Tue, 7 May 2019 23:08:02 -0700
Subject: [theory-seminar] TOCA-SV May 24: student talks
Message-ID:
Hi theory folks,
Our next TOCA-SV meeting will take place on May 24 at Google. Please hold
the date. The program will be announced soon.
As always, we will feature short student talks. Please reply to this email
by May 14 if you want to give a talk (ideally with a title).
Thanks,
Moses
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
From ccanonne at cs.stanford.edu Wed May 8 16:28:24 2019
From: ccanonne at cs.stanford.edu (=?UTF-8?Q?Cl=c3=a9ment_Canonne?=)
Date: Wed, 8 May 2019 16:28:24 -0700
Subject: [theory-seminar] TCS+ talk: Wednesday, May 15th, Ewin Tang,
University of Washington
Message-ID:
Hi all,
Next Wednesday (May 15th) at 10:00 AM, Ewin Tang (University of
Washington) will give a TCS+ on "Quantum-inspired classical linear
algebra algorithms: why and how?" (abstract below).
I've reserved a spot in the hangouts, so that we can watch the talk live
and ask her questions. There will be breakfast (classical, not quantum)!
See you there,
Cl?ment
-------------------------------
Speaker: Ewin Tang (University of Washington)
Title: Quantum-inspired classical linear algebra algorithms: why and how?
Abstract: Over the past ten years, the field of quantum machine
learning (QML) has produced many polylogarithmic-time procedures for
linear algebra routines, assuming certain "state preparation"
assumptions. Though such algorithms are formally incomparable with
classical computing, a recent line of work uses an analogous classical
model of computation as an effective point of comparison to reveal
speedups (or lack thereof) gained by QML. The resulting "dequantized"
algorithms assume sampling access to input to speed up runtimes to
polylogarithmic in input size.
In this talk, we will discuss the motivation behind this model and its
relation to existing randomized linear algebra literature. Then, we will
delve into an example quantum-inspired algorithm: Gilyen, Lloyd, and
Tang's algorithm for low-rank matrix inversion. This dequantizes a
variant of Harrow, Hassidim, and Lloyd's matrix inversion algorithm, a
seminal work in QML. Finally, we will consider the implications of this
work on exponential speedups in QML. No background of quantum computing
is assumed for this talk.
From ccanonne at cs.stanford.edu Wed May 8 17:34:11 2019
From: ccanonne at cs.stanford.edu (=?UTF-8?Q?Cl=c3=a9ment_Canonne?=)
Date: Wed, 8 May 2019 17:34:11 -0700
Subject: [theory-seminar] Theory Happy Hour: S02E04 ("Crossover Episode")
Message-ID:
Hi everyone,
It's very nice outside, I can definitely see something shining in the
sky from my office if I don't carefully block the window. That seems
like a good time to hang out, in- or outside!
And if the shiny-object-up-high justification wasn't enough, we have a
visitor this week: Ryan O'Donnell, crossing the country from the East
coast, will be around.
That amounts to strictly more than one reason to have a *happy hour*,
this *Friday* at *4:30pm* (just after the theory seminar). We'll convene
on the 4th floor (463A, as usual) and will proceed from there to enjoy
food, beer, wine, and an lack of impending deadlines. If weather allows
(as it should), we will head outwards to find a nice spot on campus.
See you on Friday!
PS: as usual, if you have preferences or restrictions on the
food/drinks, please send me an email so that I can accommodate them.
-- Cl?ment
From ofirgeri at stanford.edu Fri May 10 12:04:59 2019
From: ofirgeri at stanford.edu (Ofir Geri)
Date: Fri, 10 May 2019 19:04:59 +0000
Subject: [theory-seminar] Theory Seminar (5/10): Kira Goldner
In-Reply-To:
References:
Message-ID:
Reminder: Kira's talk is today at 3pm.
________________________________
From: Ofir Geri
Sent: Monday, May 6, 2019 3:37 PM
To: thseminar at cs.stanford.edu
Subject: Theory Seminar (5/10): Kira Goldner
Hi all,
This week in the theory seminar, Kira Goldner (University of Washington) will give a talk on Mechanism Design under Interdependent Valuations (see abstract below). The talk will be as usual on Friday, 3:00pm in Gates 463A.
The abstracts of past and upcoming seminar talks is also available on the theory seminar webpage:
http://theory.stanford.edu/seminar/
Hope to see you there!
Ofir
Mechanism Design under Interdependent Valuations
Speaker: Kira Goldner (University of Washington)
We study buyers with interdependent valuations: where one buyer's private information about an item impacts how much another buyer is willing to pay for it. In this setting, if a buyer misreports his own private information, he can impact not only what the auctioneer believes his own value is, but also what the auctioneer believes regarding other buyers' values as well, allowing him to essentially corrupt their values. As a result, the usual mechanism design tricks fall short, and welfare maximization is notoriously difficult. Almost all prior work in this setting requires a very strong "single-crossing" condition on the valuation functions in order to obtain any positive results.
We introduce two more natural notions -- first, a relaxed, parameterized notion of single-crossing, and second, a completely unrelated notion, one of submodularity over the private information of buyers -- that each separately enable good approximation guarantees to optimal welfare. These conditions, combined with the lens of approximation, allow us to go beyond not only the restrictive single-crossing notion, but to go far beyond the single-item setting and to give positive results for even the very general setting of combinatorial auctions.
Joint work with Alon Eden, Michal Feldman, Amos Fiat, and Anna Karlin.
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
From ccanonne at stanford.edu Fri May 10 12:08:40 2019
From: ccanonne at stanford.edu (Clement Louis Arthur Canonne)
Date: Fri, 10 May 2019 19:08:40 +0000
Subject: [theory-seminar] Theory Happy Hour: S02E04 ("Crossover Episode")
In-Reply-To:
References:
Message-ID:
Reminder: the post-seminar happy hour is today (after the seminar).
I can promise you the End of the World.*
* For Qu?becquois.
-- Cl?ment
On May 8, 2019 5:34 PM, Cl?ment Canonne wrote:
Hi everyone,
It's very nice outside, I can definitely see something shining in the
sky from my office if I don't carefully block the window. That seems
like a good time to hang out, in- or outside!
And if the shiny-object-up-high justification wasn't enough, we have a
visitor this week: Ryan O'Donnell, crossing the country from the East
coast, will be around.
That amounts to strictly more than one reason to have a *happy hour*,
this *Friday* at *4:30pm* (just after the theory seminar). We'll convene
on the 4th floor (463A, as usual) and will proceed from there to enjoy
food, beer, wine, and an lack of impending deadlines. If weather allows
(as it should), we will head outwards to find a nice spot on campus.
See you on Friday!
PS: as usual, if you have preferences or restrictions on the
food/drinks, please send me an email so that I can accommodate them.
-- Cl?ment
_______________________________________________
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 ccanonne at stanford.edu Fri May 10 16:00:09 2019
From: ccanonne at stanford.edu (Clement Louis Arthur Canonne)
Date: Fri, 10 May 2019 23:00:09 +0000
Subject: [theory-seminar] Theory Happy Hour: S02E04 ("Crossover Episode")
In-Reply-To:
References: ,
Message-ID: <267ffd81-dfbf-4ae8-851f-b36b3624151b@email.android.com>
So, the plan is to go there, in this newly open street next to Gates.
-- Cl?ment
On May 10, 2019 12:08 PM, Clement Louis Arthur Canonne wrote:
Reminder: the post-seminar happy hour is today (after the seminar).
I can promise you the End of the World.*
* For Qu?becquois.
-- Cl?ment
On May 8, 2019 5:34 PM, Cl?ment Canonne wrote:
Hi everyone,
It's very nice outside, I can definitely see something shining in the
sky from my office if I don't carefully block the window. That seems
like a good time to hang out, in- or outside!
And if the shiny-object-up-high justification wasn't enough, we have a
visitor this week: Ryan O'Donnell, crossing the country from the East
coast, will be around.
That amounts to strictly more than one reason to have a *happy hour*,
this *Friday* at *4:30pm* (just after the theory seminar). We'll convene
on the 4th floor (463A, as usual) and will proceed from there to enjoy
food, beer, wine, and an lack of impending deadlines. If weather allows
(as it should), we will head outwards to find a nice spot on campus.
See you on Friday!
PS: as usual, if you have preferences or restrictions on the
food/drinks, please send me an email so that I can accommodate them.
-- Cl?ment
_______________________________________________
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:
-------------- next part --------------
A non-text attachment was scrubbed...
Name: 20190510_155811.jpg
Type: image/jpeg
Size: 1956029 bytes
Desc: 20190510_155811.jpg
URL:
From ccanonne at stanford.edu Fri May 10 16:18:33 2019
From: ccanonne at stanford.edu (Clement Louis Arthur Canonne)
Date: Fri, 10 May 2019 23:18:33 +0000
Subject: [theory-seminar] Theory Happy Hour: S02E04 ("Crossover Episode")
In-Reply-To: <267ffd81-dfbf-4ae8-851f-b36b3624151b@email.android.com>
References: ,
,
<267ffd81-dfbf-4ae8-851f-b36b3624151b@email.android.com>
Message-ID:
Happening now! The spot is between Annette's Caf? and Stauffer II, next to Gates and the Gilbert Biological Sciences building.
-- Cl?ment
On May 10, 2019 4:00 PM, Clement Louis Arthur Canonne wrote:
So, the plan is to go there, in this newly open street next to Gates.
-- Cl?ment
On May 10, 2019 12:08 PM, Clement Louis Arthur Canonne wrote:
Reminder: the post-seminar happy hour is today (after the seminar).
I can promise you the End of the World.*
* For Qu?becquois.
-- Cl?ment
On May 8, 2019 5:34 PM, Cl?ment Canonne wrote:
Hi everyone,
It's very nice outside, I can definitely see something shining in the
sky from my office if I don't carefully block the window. That seems
like a good time to hang out, in- or outside!
And if the shiny-object-up-high justification wasn't enough, we have a
visitor this week: Ryan O'Donnell, crossing the country from the East
coast, will be around.
That amounts to strictly more than one reason to have a *happy hour*,
this *Friday* at *4:30pm* (just after the theory seminar). We'll convene
on the 4th floor (463A, as usual) and will proceed from there to enjoy
food, beer, wine, and an lack of impending deadlines. If weather allows
(as it should), we will head outwards to find a nice spot on campus.
See you on Friday!
PS: as usual, if you have preferences or restrictions on the
food/drinks, please send me an email so that I can accommodate them.
-- Cl?ment
_______________________________________________
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 ofirgeri at stanford.edu Mon May 13 07:44:40 2019
From: ofirgeri at stanford.edu (Ofir Geri)
Date: Mon, 13 May 2019 14:44:40 +0000
Subject: [theory-seminar] Theory Seminar (5/17): Amir Yehudayoff
Message-ID:
Hi all,
This week on Friday, Amir Yehudayoff (Technion) will give a theory seminar talk on Anti-Concentration in Most Directions (see abstract below). The talk will be as usual at 3:00pm in Gates 463A.
The abstracts of past and upcoming seminar talks are also available on the theory seminar webpage:
http://theory.stanford.edu/seminar/
Hope to see you there!
Ofir
Anti-Concentration in Most Directions
Speaker: Amir Yehudayoff (Technion)
The talk is about anti-concentration of the inner product of two independent random vectors in Euclidean space. We shall discuss a proof, and some applications.
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
From shivamgarg at stanford.edu Mon May 13 11:30:06 2019
From: shivamgarg at stanford.edu (Shivam Garg)
Date: Mon, 13 May 2019 18:30:06 +0000
Subject: [theory-seminar] Algorithms and Friends Lunch,
Monday (May 13) at Gates 463A
In-Reply-To:
References:
Message-ID:
Reminder, this is starting in 30 mins!
________________________________
From: Shivam Garg
Sent: Tuesday, May 7, 2019 8:31:45 PM
To: algorithms-and-friends at lists.stanford.edu
Cc: Avidit Raj Acharya; Megan D. Harris
Subject: Algorithms and Friends Lunch, Monday (May 13) at Gates 463A
Hi everyone,
Avidit Acharya will be giving a talk on coming Monday (May 13th), at noon, in Gates 463A.
Title: Matching Refugees to Host Country Locations Based on Preferences and Outcomes (with Kirk Bansak and Jens Hainmueller)
Abstract: Facilitating the integration of refugees has become a major policy challenge in many host countries in the context of the global displacement crisis. One of the first policy decisions host countries make in the resettlement process is the assignment of refugees to locations within the country. We develop a mechanism to match refugees to locations in a way that takes into account their expected integration outcomes and their preferences over where to be settled. Our proposal is based on a priority mechanism that allows the government first to specify a threshold g for the minimum level of expected integration success that should be achieved. Refugees are then matched to locations based on their preferences subject to meeting the government's specified threshold. The mechanism is both strategy-proof and constrained efficient in that it always generates a matching that is not Pareto dominated by any other matching that respects the government's threshold. We demonstrate our approach using simulations and a real-world application to refugee data from the United States.
Speaker Bio: Avidit (Avi) Acharya is an assistant professor of political science at Stanford. He is a formal political theorist and political economist whose work ranges across a diverse set of topics including voting theory, bargaining theory, principal-agent theory, behavioral political economy, distributive politics, and long run development. Avi?s papers have been published (or are forthcoming) in the leading journals of political science, including the American Political Science Review, American Journal of Political Science, and Journal of Politics, as well as the top journals in economic theory, including Econometrica, Journal of Economic Theory, and Games and Economic Behavior. Avi earned his PhD in political economy from Princeton University, and taught in the political science and economics departments of the University of Rochester before coming to Stanford.
Thanks,
Shivam
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
From wyma at stanford.edu Mon May 13 14:42:07 2019
From: wyma at stanford.edu (Weiyun Ma)
Date: Mon, 13 May 2019 21:42:07 +0000
Subject: [theory-seminar] Theory Lunch 5/16 -- Joshua Brakensiek
Message-ID:
Hi all,
This Thursday at theory lunch, Josh will tell us about "Bridging between 0/1 and Linear Programming via Random Walks." (See abstract below.)
As usual, please join us from noon to 1pm at 463A.
----------------------------------------------------------
Bridging between 0/1 and Linear Programming via Random Walks
Speaker: Joshua Brakensiek
Under the Strong Exponential Time Hypothesis, an integer linear program with n Boolean-valued variables and m equations cannot be solved in c^n time for any constant c < 2. If the domain of the variables is relaxed to [0,1], the associated linear program can of course be solved in polynomial time. In this work, we give a natural algorithmic bridging between these extremes of 0-1 and linear programming. Specifically, for any subset (finite union of intervals) E \subset [0,1] containing {0, 1}, we give a random-walk based algorithm with runtime ~(2-measure(E))^n (up to polynomial factors) that finds a solution in E^n to any n-variable linear program with m constraints that is feasible over {0,1}^n. Note that as E expands from {0,1} to [0,1], the runtime improves smoothly from 2^n to polynomial.
Taking E = [0,1/k) \cup (1-1/k,1] in our result yields as a corollary a randomized ~(2-2/k)^n time algorithm for
k-SAT, recovering a result of Schoning (FOCS 1999). While our approach has some high level resemblance to Schoning's beautiful algorithm, our general algorithm is based on a more sophisticated random walk that incorporates several new ingredients, such as a multiplicative potential to measure progress, a judicious choice of starting distribution, and a time varying distribution for the evolution of the random walk that is itself computed via an LP at each step (a solution to which is guaranteed based on the minimax theorem).
Joint work with Venkatesan Guruswami (CMU).
----------------------------------------------------------
Best,
Anna
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
From wyma at stanford.edu Mon May 13 14:42:07 2019
From: wyma at stanford.edu (Weiyun Ma)
Date: Mon, 13 May 2019 21:42:07 +0000
Subject: [theory-seminar] Theory Lunch 5/16 -- Joshua Brakensiek
Message-ID:
Hi all,
This Thursday at theory lunch, Josh will tell us about "Bridging between 0/1 and Linear Programming via Random Walks." (See abstract below.)
As usual, please join us from noon to 1pm at 463A.
----------------------------------------------------------
Bridging between 0/1 and Linear Programming via Random Walks
Speaker: Joshua Brakensiek
Under the Strong Exponential Time Hypothesis, an integer linear program with n Boolean-valued variables and m equations cannot be solved in c^n time for any constant c < 2. If the domain of the variables is relaxed to [0,1], the associated linear program can of course be solved in polynomial time. In this work, we give a natural algorithmic bridging between these extremes of 0-1 and linear programming. Specifically, for any subset (finite union of intervals) E \subset [0,1] containing {0, 1}, we give a random-walk based algorithm with runtime ~(2-measure(E))^n (up to polynomial factors) that finds a solution in E^n to any n-variable linear program with m constraints that is feasible over {0,1}^n. Note that as E expands from {0,1} to [0,1], the runtime improves smoothly from 2^n to polynomial.
Taking E = [0,1/k) \cup (1-1/k,1] in our result yields as a corollary a randomized ~(2-2/k)^n time algorithm for
k-SAT, recovering a result of Schoning (FOCS 1999). While our approach has some high level resemblance to Schoning's beautiful algorithm, our general algorithm is based on a more sophisticated random walk that incorporates several new ingredients, such as a multiplicative potential to measure progress, a judicious choice of starting distribution, and a time varying distribution for the evolution of the random walk that is itself computed via an LP at each step (a solution to which is guaranteed based on the minimax theorem).
Joint work with Venkatesan Guruswami (CMU).
----------------------------------------------------------
Best,
Anna
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
From ccanonne at stanford.edu Tue May 14 20:26:11 2019
From: ccanonne at stanford.edu (Clement Louis Arthur Canonne)
Date: Wed, 15 May 2019 03:26:11 +0000
Subject: [theory-seminar] TCS+ talk: Wednesday, May 15th, Ewin Tang,
University of Washington
In-Reply-To:
References:
Message-ID:
Reminder: this is tomorrow!
-- Cl?ment
On May 8, 2019 4:28 PM, Cl?ment Canonne wrote:
Hi all,
Next Wednesday (May 15th) at 10:00 AM, Ewin Tang (University of
Washington) will give a TCS+ on "Quantum-inspired classical linear
algebra algorithms: why and how?" (abstract below).
I've reserved a spot in the hangouts, so that we can watch the talk live
and ask her questions. There will be breakfast (classical, not quantum)!
See you there,
Cl?ment
-------------------------------
Speaker: Ewin Tang (University of Washington)
Title: Quantum-inspired classical linear algebra algorithms: why and how?
Abstract: Over the past ten years, the field of quantum machine
learning (QML) has produced many polylogarithmic-time procedures for
linear algebra routines, assuming certain "state preparation"
assumptions. Though such algorithms are formally incomparable with
classical computing, a recent line of work uses an analogous classical
model of computation as an effective point of comparison to reveal
speedups (or lack thereof) gained by QML. The resulting "dequantized"
algorithms assume sampling access to input to speed up runtimes to
polylogarithmic in input size.
In this talk, we will discuss the motivation behind this model and its
relation to existing randomized linear algebra literature. Then, we will
delve into an example quantum-inspired algorithm: Gilyen, Lloyd, and
Tang's algorithm for low-rank matrix inversion. This dequantizes a
variant of Harrow, Hassidim, and Lloyd's matrix inversion algorithm, a
seminal work in QML. Finally, we will consider the implications of this
work on exponential speedups in QML. No background of quantum computing
is assumed for this talk.
_______________________________________________
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 ofirgeri at stanford.edu Thu May 16 09:23:06 2019
From: ofirgeri at stanford.edu (Ofir Geri)
Date: Thu, 16 May 2019 16:23:06 +0000
Subject: [theory-seminar] Talk Today at 4:30pm: Ron Lavi
In-Reply-To:
References: ,
Message-ID:
Hi all,
Ron Lavi is a giving a talk today that may be of interest to some of you (see details below).
Best,
Ofir
---------- Forwarded message ---------
Speaker: Ron Lavi
Location: Spilker 317
Time: Thursday March 16th, at 4:30
Title: Contests for Revenue Share
Abstract: In a framework of contest design, we study capacity-constrained competition and analyze its resulting revenue shares. Our analysis contrasts the near-symmetric case, where firms have similar supply sizes, and the extremely asymmetric case, where one large firm dominates the market. In particular, we show that while in the near-symmetric case simple contests provide near optimal equilibrium revenues for all sellers, in the asymmetric case a large firm can design more complicated contests that result in disproportionally low equilibrium revenues for her smaller opponents.
This is joint work with Uriel Feige and Moshe Tennenholtz.
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
From moses at cs.stanford.edu Thu May 16 13:50:17 2019
From: moses at cs.stanford.edu (Moses Charikar)
Date: Thu, 16 May 2019 13:50:17 -0700
Subject: [theory-seminar] 6th TOCA SV Day Announcement
In-Reply-To:
References:
Message-ID:
The 6th edition of TOCA SV (Silicon Valley Theory Day) will be held in the
Google Mountain View Campus on Friday, May 24, 2019.
For more details, please see the web page
.
If you intend to attend, please complete the short registration included there
(this is REQUIRED to make sure we can prepare visitor badges to enter the
Google Building). Please also feel welcome to forward this message to CS
theorists in the Bay Area who may be interested.
Thanks, and look forward to seeing you at Google on May 24th!
Moses Charikar (Stanford)
Ravi Kumar (Google Research)
Omer Reingold (Stanford)
D. Sivakumar (Google Research)
Greg Valiant (Stanford)
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
From ofirgeri at stanford.edu Fri May 17 12:43:03 2019
From: ofirgeri at stanford.edu (Ofir Geri)
Date: Fri, 17 May 2019 19:43:03 +0000
Subject: [theory-seminar] Theory Seminar (5/17): Amir Yehudayoff
In-Reply-To:
References:
Message-ID:
Reminder: Amir's talk is today at 3pm.
________________________________
From: Ofir Geri
Sent: Monday, May 13, 2019 7:44 AM
To: thseminar at cs.stanford.edu
Subject: Theory Seminar (5/17): Amir Yehudayoff
Hi all,
This week on Friday, Amir Yehudayoff (Technion) will give a theory seminar talk on Anti-Concentration in Most Directions (see abstract below). The talk will be as usual at 3:00pm in Gates 463A.
The abstracts of past and upcoming seminar talks are also available on the theory seminar webpage:
http://theory.stanford.edu/seminar/
Hope to see you there!
Ofir
Anti-Concentration in Most Directions
Speaker: Amir Yehudayoff (Technion)
The talk is about anti-concentration of the inner product of two independent random vectors in Euclidean space. We shall discuss a proof, and some applications.
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
From shivamgarg at stanford.edu Sat May 18 10:22:38 2019
From: shivamgarg at stanford.edu (Shivam Garg)
Date: Sat, 18 May 2019 17:22:38 +0000
Subject: [theory-seminar] Algorithms and Friends Lunch, Monday (May 20),
in Gates 463A
Message-ID:
Hi everyone,
Hanlee P. Ji will be giving a talk on coming Monday (May 20th), at noon, in Gates 463A.
Title: Characterizing structural variation in human genome code
Abstract: Structural variants (SVs) involve complex rearrangements of the human genome code, occur throughout the human genome, and may encompass alterations spanning Megabases of DNA sequences. These rearrangements lead to a variety of genetic disorders that include cancer, congenital malformations, and neurocognitive conditions such as autism. The detection and characterization of germline and somatic rearrangements are critical for determining the genetic pathogenesis of these diseases and provide valuable diagnostic information. The majority of genomic studies characterizing SVs rely on short sequence reads (
From shivamgarg at stanford.edu Mon May 20 09:44:59 2019
From: shivamgarg at stanford.edu (Shivam Garg)
Date: Mon, 20 May 2019 16:44:59 +0000
Subject: [theory-seminar] Algorithms and Friends Lunch, Monday (May 20),
in Gates 463A
In-Reply-To:
References:
Message-ID:
Reminder, this is happening today at noon.
________________________________
From: Shivam Garg
Sent: Saturday, May 18, 2019 10:22:38 AM
To: algorithms-and-friends at lists.stanford.edu; theory-seminar at lists.stanford.edu
Cc: Hanlee P. Ji; Megan D. Harris
Subject: Algorithms and Friends Lunch, Monday (May 20), in Gates 463A
Hi everyone,
Hanlee P. Ji will be giving a talk on coming Monday (May 20th), at noon, in Gates 463A.
Title: Characterizing structural variation in human genome code
Abstract: Structural variants (SVs) involve complex rearrangements of the human genome code, occur throughout the human genome, and may encompass alterations spanning Megabases of DNA sequences. These rearrangements lead to a variety of genetic disorders that include cancer, congenital malformations, and neurocognitive conditions such as autism. The detection and characterization of germline and somatic rearrangements are critical for determining the genetic pathogenesis of these diseases and provide valuable diagnostic information. The majority of genomic studies characterizing SVs rely on short sequence reads (
From wyma at stanford.edu Mon May 20 13:23:51 2019
From: wyma at stanford.edu (Weiyun Ma)
Date: Mon, 20 May 2019 20:23:51 +0000
Subject: [theory-seminar] Theory Lunch 5/23 -- Nitya Mani
Message-ID:
Hi everyone,
This Thursday at theory lunch, Nitya will tell us about "Max-k-Cuts in H-Free Graphs." (See abstract below.)
As always, please join us from noon to 1pm at 463A.
----------------------------------------------------------
Max-k-Cuts in H-Free Graphs
Speaker: Nitya Mani
We study the following question: how few edges can we delete from any H-free graph on n vertices in order to make the resulting graph k-colorable? It turns out that various classical problems in extremal combinatorics are special cases of this problem, and this quantity yields bounds on the Max-k-Cut of an H-free graph. For H any fixed odd cycle, we determine the answer up to a constant factor when n is sufficiently large. We also prove upper bounds when H is a clique or lies in some other families of graphs.
----------------------------------------------------------
Best,
Anna
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
From wyma at stanford.edu Mon May 20 13:23:51 2019
From: wyma at stanford.edu (Weiyun Ma)
Date: Mon, 20 May 2019 20:23:51 +0000
Subject: [theory-seminar] Theory Lunch 5/23 -- Nitya Mani
Message-ID:
Hi everyone,
This Thursday at theory lunch, Nitya will tell us about "Max-k-Cuts in H-Free Graphs." (See abstract below.)
As always, please join us from noon to 1pm at 463A.
----------------------------------------------------------
Max-k-Cuts in H-Free Graphs
Speaker: Nitya Mani
We study the following question: how few edges can we delete from any H-free graph on n vertices in order to make the resulting graph k-colorable? It turns out that various classical problems in extremal combinatorics are special cases of this problem, and this quantity yields bounds on the Max-k-Cut of an H-free graph. For H any fixed odd cycle, we determine the answer up to a constant factor when n is sufficiently large. We also prove upper bounds when H is a clique or lies in some other families of graphs.
----------------------------------------------------------
Best,
Anna
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
From ofirgeri at stanford.edu Mon May 20 13:34:22 2019
From: ofirgeri at stanford.edu (Ofir Geri)
Date: Mon, 20 May 2019 20:34:22 +0000
Subject: [theory-seminar] Theory Seminar (Thursday 5/23): Andrea Lincoln
Message-ID:
Hi all,
This week, Andrea Lincoln (MIT) will give a theory seminar talk on Faster Random k-CNF Satisfiability (see abstract below).
The talk will be on THURSDAY at 4:15 PM in Gates 463A. Please note the non-standard day and time.
The abstracts of past and upcoming seminar talks are also available on the theory seminar webpage:
http://theory.stanford.edu/seminar/
Hope to see you there!
Ofir
Faster Random k-CNF Satisfiability
Speaker: Andrea Lincoln (MIT)
We describe an algorithm to solve the problem of Boolean CNF-Satisfiability when the input formula is chosen randomly.
We build upon the algorithms of Sch?ning 1999 and Dantsin et al. in 2002. The Sch?ning algorithm works by trying many possible random assignments, and for each one searching systematically in the neighborhood of that assignment for a satisfying solution. Previous algorithms for this problem run in time O(2^{n (1- \Omega(1)/k)}).
Our improvement is simple: we count how many clauses are satisfied by each randomly sampled assignment, and only search in the neighborhoods of assignments with abnormally many satisfied clauses. We show that assignments like these are significantly more likely to be near a satisfying assignment. This improvement saves a factor of 2^{n \Omega(\lg^2 k)/k}, resulting in an overall runtime of O(2^{n (1- \Omega(\lg^2 k)/k)}) for random k-SAT.
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
From ofirgeri at stanford.edu Thu May 23 10:00:31 2019
From: ofirgeri at stanford.edu (Ofir Geri)
Date: Thu, 23 May 2019 17:00:31 +0000
Subject: [theory-seminar] Theory Seminar (Thursday 5/23): Andrea Lincoln
In-Reply-To:
References:
Message-ID:
Reminder: Andrea's talk is today is 4:15pm.
________________________________
From: Ofir Geri
Sent: Monday, May 20, 2019 1:34 PM
To: thseminar at cs.stanford.edu
Subject: Theory Seminar (Thursday 5/23): Andrea Lincoln
Hi all,
This week, Andrea Lincoln (MIT) will give a theory seminar talk on Faster Random k-CNF Satisfiability (see abstract below).
The talk will be on THURSDAY at 4:15 PM in Gates 463A. Please note the non-standard day and time.
The abstracts of past and upcoming seminar talks are also available on the theory seminar webpage:
http://theory.stanford.edu/seminar/
Hope to see you there!
Ofir
Faster Random k-CNF Satisfiability
Speaker: Andrea Lincoln (MIT)
We describe an algorithm to solve the problem of Boolean CNF-Satisfiability when the input formula is chosen randomly.
We build upon the algorithms of Sch?ning 1999 and Dantsin et al. in 2002. The Sch?ning algorithm works by trying many possible random assignments, and for each one searching systematically in the neighborhood of that assignment for a satisfying solution. Previous algorithms for this problem run in time O(2^{n (1- \Omega(1)/k)}).
Our improvement is simple: we count how many clauses are satisfied by each randomly sampled assignment, and only search in the neighborhoods of assignments with abnormally many satisfied clauses. We show that assignments like these are significantly more likely to be near a satisfying assignment. This improvement saves a factor of 2^{n \Omega(\lg^2 k)/k}, resulting in an overall runtime of O(2^{n (1- \Omega(\lg^2 k)/k)}) for random k-SAT.
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
From psimin at stanford.edu Fri May 24 09:02:00 2019
From: psimin at stanford.edu (Paris Syminelakis)
Date: Fri, 24 May 2019 16:02:00 +0000
Subject: [theory-seminar] Paris's Thesis Defense May 30th, 3pm, Gates 104
Message-ID:
Hi theory folks,
I would like to invite you to my thesis defense taking place next Thursday May 30th, 3pm at Gates 104. Details follow.
Hope to see you there!
-- Paris
University Ph.D. Dissertation Defense
Department of Electrical Engineering
Paris Syminelakis
Kernel Evaluation in High Dimensions:
Importance Sampling and Nearest Neighbor Search
Advisor: Prof. Moses Charikar
Date: Thursday, May 30, 2019
Time: 3:00 pm (refreshments at 2:45 pm)
Location: Gates 104
High-dimensional datasets arise in a multitude of ways: naturally, due to the need to represent complex objects (images, audio), and synthetically, due to the need to represent complex relationships between objects through vector embeddings (semantic relationships e.g. GloVe, class labels e.g. ImageNet). At the same time, these datasets are of massive scale on the order of million points and even executing simple computational tasks, like evaluating a predictive model, for all points in the dataset can be time consuming. These considerations along with the fact that such computational primitives are building blocks for applications that are becoming ubiquitous in everyday life raise the need for efficient and reliable algorithms.
In this thesis, we study the problem of fast kernel evaluation that underlies all kernel-based predictive models. Given a kernel function k(x,y) that takes as input two vectors and outputs a number in [0,1], we ask the question of how fast we can output an approximation to the average value of the kernel function (kernel density) between a query vector q and all points in a dataset. This problem has a long history of study but before our work under worst case assumptions all methods either run in time exponential in the dimension or linear in the number of points in the dataset. We develop a new set of techniques based on randomized space partitions that result in algorithms that can provably approximate the kernel density in sub-linear time for a wide range of kernel functions. One of the main themes of our work is designing low-variance importance sampling schemes through hashing. We complement our positive algorithmic results with impossibility results showing that obtaining significantly faster algorithms for this problem will either refute popular computational complexity conjectures or result in faster algorithms for nearest neighbor search.
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
From ofirgeri at stanford.edu Tue May 28 13:33:06 2019
From: ofirgeri at stanford.edu (Ofir Geri)
Date: Tue, 28 May 2019 20:33:06 +0000
Subject: [theory-seminar] Theory Seminar (5/31): Michal Moshkovitz
Message-ID:
Hi all,
This week Michal Moshkovitz (UCSD) will give a theory seminar talk: On Bounded-Memory Learning (see abstract below). The talk will be as usual on Friday 3:00 PM in Gates 463A.
The abstracts of past and upcoming seminar talks are also available on the theory seminar webpage:
http://theory.stanford.edu/seminar/
Hope to see you there!
Ofir
On Bounded-Memory Learning
Speaker: Michal Moshkovitz (UCSD)
One can learn any hypothesis class H with O(log |H|) labeled examples. Alas, learning with so few examples requires saving the examples in memory, and this requires |X|^(O(log|H|)) memory states, where X is the set of all labeled examples. This motivates the question of how many labeled examples are needed in case the memory is bounded. One might wonder whether a general combinatorial condition exists for (un)learnability with bounded memory. In this talk we give a combinatorial condition for learnability with bounded memory and a combinatorial condition for unlearnability with bounded memory.
The talk is based on joint works with Dana Moshkovitz and Naftali Tishby.
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
From wyma at stanford.edu Tue May 28 13:57:08 2019
From: wyma at stanford.edu (Weiyun Ma)
Date: Tue, 28 May 2019 20:57:08 +0000
Subject: [theory-seminar] Theory lunch this week
Message-ID:
Hi all,
Unfortunately our speaker this Thursday may have to cancel due to sickness. If you would like to give a talk or do activities with theory folks, please let me know.
Best,
Anna
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
From wyma at stanford.edu Tue May 28 13:57:08 2019
From: wyma at stanford.edu (Weiyun Ma)
Date: Tue, 28 May 2019 20:57:08 +0000
Subject: [theory-seminar] Theory lunch this week
Message-ID:
Hi all,
Unfortunately our speaker this Thursday may have to cancel due to sickness. If you would like to give a talk or do activities with theory folks, please let me know.
Best,
Anna
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
From wyma at stanford.edu Tue May 28 22:05:33 2019
From: wyma at stanford.edu (Weiyun Ma)
Date: Wed, 29 May 2019 05:05:33 +0000
Subject: [theory-seminar] Theory Lunch 5/30 -- Michel Schellekens
Message-ID:
Hi all,
Michel Schellekens has kindly offered to speak at theory lunch this Thursday. He will tell us about "Modular Algorithm Analysis *by example*." (See abstract below.)
As always, please join us from noon to 1pm at 463A.
----------------------------------------------------------
Modular Algorithm Analysis *by example*
Speaker: Michel Schellekens
Timing modularity guarantees that the running time of an algorithm is a combination of the times of its parts. The property, when it holds, drastically simplifies timing, supporting recurrence equation extraction for worst-case time, average case time and second moments. We focus on modularity of the average-case time for comparison-based algorithms that gather information by comparing data. The Modular Quantitative Analysis framework (MOQA) provides a mathematical model of computation for the modular analysis of comparison-based algorithms.
The model of computation consists of: (a) modelling data structures as partially-ordered finite sets; (b) modelling data on these by topological sorts; (c) considering computation states as finite multisets of such data; (d) analysing algorithms by their induced transformations on states. In this view, an abstract specification of a sorting algorithm has input state given by any possible permutation of a finite set of elements (represented, according to (a) and (b), by a discrete partially-ordered set together with its topological sorts given by all permutations) and output state a sorted list of elements (represented, again according to (a) and (b), by a linearly-ordered finite set with its unique topological sort).
Series-parallel (SP) orders form an important, and computationally tractable, class of data structures. These include trees and play a role in sorting, sequencing, and scheduling. We focus on the least MOQA fragment sufficient to support modular time analysis over SP-orders. The fragment is based on the ?diffusion operation? which generalizes well-known data structure operations, including the heap creation operation and insertion sort?s insert operation. Both rely on repeated push-ups or push-downs that ?diffuse? data (i.e. topological sorts) of a pair of data structures over a larger whole. The final data must respect the newly created order, i.e. form a new topological sort?which is guaranteed for the diffusion operation. The talk introduces the main ideas of modular algorithmic analysis by example and the recurrence equations involved.
----------------------------------------------------------
Best,
Anna
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
From wyma at stanford.edu Tue May 28 22:05:33 2019
From: wyma at stanford.edu (Weiyun Ma)
Date: Wed, 29 May 2019 05:05:33 +0000
Subject: [theory-seminar] Theory Lunch 5/30 -- Michel Schellekens
Message-ID:
Hi all,
Michel Schellekens has kindly offered to speak at theory lunch this Thursday. He will tell us about "Modular Algorithm Analysis *by example*." (See abstract below.)
As always, please join us from noon to 1pm at 463A.
----------------------------------------------------------
Modular Algorithm Analysis *by example*
Speaker: Michel Schellekens
Timing modularity guarantees that the running time of an algorithm is a combination of the times of its parts. The property, when it holds, drastically simplifies timing, supporting recurrence equation extraction for worst-case time, average case time and second moments. We focus on modularity of the average-case time for comparison-based algorithms that gather information by comparing data. The Modular Quantitative Analysis framework (MOQA) provides a mathematical model of computation for the modular analysis of comparison-based algorithms.
The model of computation consists of: (a) modelling data structures as partially-ordered finite sets; (b) modelling data on these by topological sorts; (c) considering computation states as finite multisets of such data; (d) analysing algorithms by their induced transformations on states. In this view, an abstract specification of a sorting algorithm has input state given by any possible permutation of a finite set of elements (represented, according to (a) and (b), by a discrete partially-ordered set together with its topological sorts given by all permutations) and output state a sorted list of elements (represented, again according to (a) and (b), by a linearly-ordered finite set with its unique topological sort).
Series-parallel (SP) orders form an important, and computationally tractable, class of data structures. These include trees and play a role in sorting, sequencing, and scheduling. We focus on the least MOQA fragment sufficient to support modular time analysis over SP-orders. The fragment is based on the ?diffusion operation? which generalizes well-known data structure operations, including the heap creation operation and insertion sort?s insert operation. Both rely on repeated push-ups or push-downs that ?diffuse? data (i.e. topological sorts) of a pair of data structures over a larger whole. The final data must respect the newly created order, i.e. form a new topological sort?which is guaranteed for the diffusion operation. The talk introduces the main ideas of modular algorithmic analysis by example and the recurrence equations involved.
----------------------------------------------------------
Best,
Anna
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
From ofirgeri at stanford.edu Fri May 31 11:50:07 2019
From: ofirgeri at stanford.edu (Ofir Geri)
Date: Fri, 31 May 2019 18:50:07 +0000
Subject: [theory-seminar] Theory Seminar (5/31): Michal Moshkovitz
In-Reply-To:
References:
Message-ID:
Reminder: Michal's talk is today at 3pm.
________________________________
From: Ofir Geri
Sent: Tuesday, May 28, 2019 1:33 PM
To: thseminar at cs.stanford.edu
Subject: Theory Seminar (5/31): Michal Moshkovitz
Hi all,
This week Michal Moshkovitz (UCSD) will give a theory seminar talk: On Bounded-Memory Learning (see abstract below). The talk will be as usual on Friday 3:00 PM in Gates 463A.
The abstracts of past and upcoming seminar talks are also available on the theory seminar webpage:
http://theory.stanford.edu/seminar/
Hope to see you there!
Ofir
On Bounded-Memory Learning
Speaker: Michal Moshkovitz (UCSD)
One can learn any hypothesis class H with O(log |H|) labeled examples. Alas, learning with so few examples requires saving the examples in memory, and this requires |X|^(O(log|H|)) memory states, where X is the set of all labeled examples. This motivates the question of how many labeled examples are needed in case the memory is bounded. One might wonder whether a general combinatorial condition exists for (un)learnability with bounded memory. In this talk we give a combinatorial condition for learnability with bounded memory and a combinatorial condition for unlearnability with bounded memory.
The talk is based on joint works with Dana Moshkovitz and Naftali Tishby.
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
From hongyang at cs.stanford.edu Fri May 31 15:29:02 2019
From: hongyang at cs.stanford.edu (Hongyang Zhang)
Date: Fri, 31 May 2019 15:29:02 -0700
Subject: [theory-seminar] Hongyang's thesis defense on Monday June 3
Message-ID:
Hi friends,
I'm going to do my thesis defense on *Monday, June 3, 2019 at 10:30am*
*at Gates 463A*. You are all welcome to come! More information of the talk
can be found below.
-----------------------------------------------------------------------------------------
University Oral Examination
*Title: Algorithms and Generalization for Large-Scale Matrices and Tensors*
Hongyang Zhang
Computer Science Department
Stanford University
Advised by Ashish Goel and Gregory Valiant
*Monday, June 3, 2019 at 10:30am* (refreshments served at 10:15am)
*Gates Building, Room 463A*
*Abstract:* Over the past decade, machine learning methods such as deep
neural networks have made a huge impact on a variety of complex tasks. On
the other hand, very little is understood about when and why these ML
methods work in practice. Bridging this gap requires better understanding
of the non-convex optimization paradigm commonly used in training deep
neural networks, as well as better modeling of real world data. My thesis
aims at providing principled algorithms and insights by examining
analytically tractable objects such as matrices and tensors, that are
intimately connected to neural networks.
This talk will show a few results:
i) we study gradient based optimization methods and their generalization
performance (or sample efficiency) in over-parameterized matrix models. Our
result highlights the role of the optimization algorithm in explaining
generalization when there are more trainable parameters than the size of
the dataset.
ii) we consider the problem of predicting the missing entries of
high-dimensional tensor data. We show an interesting representation-sample
trade-off in the choice of tensor models for fitting the data.
iii) we present new methods for the classic distance query problem that
creates state of the art data structures on a variety of large-scale graph
data.
-----------------------------------------------------------------------------------------
Best,
Hongyang
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
From hongyang at cs.stanford.edu Fri May 31 15:29:02 2019
From: hongyang at cs.stanford.edu (Hongyang Zhang)
Date: Fri, 31 May 2019 15:29:02 -0700
Subject: [theory-seminar] Hongyang's thesis defense on Monday June 3
Message-ID:
Hi friends,
I'm going to do my thesis defense on *Monday, June 3, 2019 at 10:30am*
*at Gates 463A*. You are all welcome to come! More information of the talk
can be found below.
-----------------------------------------------------------------------------------------
University Oral Examination
*Title: Algorithms and Generalization for Large-Scale Matrices and Tensors*
Hongyang Zhang
Computer Science Department
Stanford University
Advised by Ashish Goel and Gregory Valiant
*Monday, June 3, 2019 at 10:30am* (refreshments served at 10:15am)
*Gates Building, Room 463A*
*Abstract:* Over the past decade, machine learning methods such as deep
neural networks have made a huge impact on a variety of complex tasks. On
the other hand, very little is understood about when and why these ML
methods work in practice. Bridging this gap requires better understanding
of the non-convex optimization paradigm commonly used in training deep
neural networks, as well as better modeling of real world data. My thesis
aims at providing principled algorithms and insights by examining
analytically tractable objects such as matrices and tensors, that are
intimately connected to neural networks.
This talk will show a few results:
i) we study gradient based optimization methods and their generalization
performance (or sample efficiency) in over-parameterized matrix models. Our
result highlights the role of the optimization algorithm in explaining
generalization when there are more trainable parameters than the size of
the dataset.
ii) we consider the problem of predicting the missing entries of
high-dimensional tensor data. We show an interesting representation-sample
trade-off in the choice of tensor models for fitting the data.
iii) we present new methods for the classic distance query problem that
creates state of the art data structures on a variety of large-scale graph
data.
-----------------------------------------------------------------------------------------
Best,
Hongyang
-------------- next part --------------
An HTML attachment was scrubbed...
URL: