From moses at cs.stanford.edu Mon Apr 1 14:02:29 2019
From: moses at cs.stanford.edu (Moses Charikar)
Date: Mon, 1 Apr 2019 14:02:29 -0700
Subject: [theory-seminar] Fwd: Probability seminar of potential interest
In-Reply-To: <223EFF04-EE70-49FA-9ACB-CD176247CD52@stanford.edu>
References:
<223EFF04-EE70-49FA-9ACB-CD176247CD52@stanford.edu>
Message-ID:
Andrea Montanari?s talk today at 4pm could be interesting to many of you.
Cheers,
Moses
---------- Forwarded message ---------
From: Andrea Montanari
Date: Mon, Apr 1, 2019 at 1:30 PM
Subject: Fwd: Probability seminar of potential interest
To: moses at cs.stanford.edu
Hi Moses,
Perhaps some cs theory students might be interested in this ...
A
--------------------
Monday, April 1, 4pm: Probability Seminar, Sequoia Hall Room 200
Andrea Montanari, Stanford University
"Optimization of the Sherrington-Kirkpatrick Hamiltonian"
Abstracts from: https://statistics.stanford.edu/events/probability-seminar
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
-------------- next part --------------
A non-text attachment was scrubbed...
Name: Apr01-2019.pdf
Type: application/pdf
Size: 100539 bytes
Desc: not available
URL:
From ofirgeri at stanford.edu Mon Apr 1 15:56:47 2019
From: ofirgeri at stanford.edu (Ofir Geri)
Date: Mon, 1 Apr 2019 22:56:47 +0000
Subject: [theory-seminar] Theory Seminar (4/5): Shant Boodaghians
Message-ID:
Hi all,
The first theory seminar of the quarter will be given by Shant Boodaghians (UIUC) on Reductions and Algorithms for Pure Nash Equilibria in Smoothed Network Coordination Games (see abstract below). The talk will be as usual on Friday, 3:00 PM in Gates 463A.
Hope to see you there!
Ofir
Reductions and Algorithms for Pure Nash Equilibria in Smoothed Network Coordination Games
Speaker: Shant Boodaghians (UIUC)
A 2-player coordination game is a game where both players receive the same payoff, and are therefore collaborating. A network coordination game is a multi-player game where the players are nodes in a graph, and each edge models a 2-player coordination game. The strategy chosen by each node applies simultaneously for all incident edges. We study the PLS-complete problem of finding a pure-strategy Nash equilibrium in such games, when the instance is smoothed. (The payoff values are independent random variables, where the distribution is chosen adversarially under a density constraint.) We show that the natural best-response dynamic finds an equilibrium in polynomial time when the number of strategies that each player may choose from is a constant.
We first introduce smoothness-preserving reductions, and show that the problem admits such a reduction to finding a locally maximal cut in a weigthed graph, which allows us to inherit recent developments in the smoothed analysis of the local-search algorithm for this problem. Note that local-max-cut is also PLS-complete. This reduction, however, only holds when each player is allowed to choose from exactly 2 strategies. In the general case, we show a reduction to finding a cut maximal up to flipping 2 nodes; this variant has not been shown to be efficient in the smoothed case, and so the reduction is conditional.
Second, we explain how the framework for the analysis of local-max-cut may be applied directly to our problem, rather than going through reductions. When players are allowed to choose from k strategies, we show that best-response searches converge in polynomial time, where the exponent is linear in k. The conditional reduction above seeks to eliminate the exponential dependence on k.
(Joint work with Rucha Kulkarni and Ruta Mehta)
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
From wyma at stanford.edu Mon Apr 1 22:19:08 2019
From: wyma at stanford.edu (Weiyun Ma)
Date: Tue, 2 Apr 2019 05:19:08 +0000
Subject: [theory-seminar] Theory Lunch 4/4 -- Open Problem Session
Message-ID:
Hi everyone,
Welcome back from spring break! Theory lunch resumes this Thursday (4/4) and we will do an open problem session. Each speaker will have 5 minutes to present an open problem. Please let me know if you?d like to take a slot (first come, first served)! As usual, we will meet from noon to 1pm at Gates 463A.
Also, don?t forget to sign up for the talks this quarter at:
https://docs.google.com/document/d/1S0QcDMTn-JRaP1cRFRihZyeNfHUmpYLkyORKMcGIBgY/edit
You can also sign up in person on Thursday. See you all soon!
Best,
Anna
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
From wyma at stanford.edu Mon Apr 1 22:19:08 2019
From: wyma at stanford.edu (Weiyun Ma)
Date: Tue, 2 Apr 2019 05:19:08 +0000
Subject: [theory-seminar] Theory Lunch 4/4 -- Open Problem Session
Message-ID:
Hi everyone,
Welcome back from spring break! Theory lunch resumes this Thursday (4/4) and we will do an open problem session. Each speaker will have 5 minutes to present an open problem. Please let me know if you?d like to take a slot (first come, first served)! As usual, we will meet from noon to 1pm at Gates 463A.
Also, don?t forget to sign up for the talks this quarter at:
https://docs.google.com/document/d/1S0QcDMTn-JRaP1cRFRihZyeNfHUmpYLkyORKMcGIBgY/edit
You can also sign up in person on Thursday. See you all soon!
Best,
Anna
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
From michael.kim at cs.stanford.edu Wed Apr 3 01:27:03 2019
From: michael.kim at cs.stanford.edu (Michael Kim)
Date: Wed, 3 Apr 2019 01:27:03 -0700
Subject: [theory-seminar] Fwd: [TheoryLunch] Fwd: Alert regarding FOCS
In-Reply-To:
References: <6c5b06c118ca442786d05837bd7185ac@IBWHBXV02.wismain.weizmann.ac.il>
Message-ID:
FYI make sure you have the right FOCS site!
---------- Forwarded message ---------
From: Irit Dinur
Date: Wed, Apr 3, 2019 at 1:15 AM
Subject: [TheoryLunch] Fwd: Alert regarding FOCS
To: TheoryLunch Mailing List
( from Yuval Rabani )
Hi everyone,
I apologize for the grand spamming. If this weren?t both urgent and
potentially serious, I wouldn?t have done it.
I?m trying to increase the exposure of the theory community to the
following message I tweeted today (and also a few days ago, unfortunately
on April 1). You?re a random (very biased) sample of theoreticians with
emails easily available to me, and I hope you can help spread this further.
Here?s the tweet (you can find it on our new account @IeeeTcmf):
????????????
As April 5 (FOCS submission deadline) approaches:
Beware of a site (aconf dot org) that pretends to be FOCS and runs a fake
submission server!
When googling "FOCS 2019", it shows up first or second (lower on bing)
Purpose unknown, phishing passwords?
????????????
I don?t know for sure that the intention is malicious, and I and others
have taken other steps regarding the potential threat, but it?s a good idea
to have as many people aware of this as possible. We (FOCS 2019 team) are
aware of at least two cases where people fell for this and noticed it. We
don?t know how many people fell for this and did not notice it.
Best and thanks for your help,
Yuval
--
--Irit
_______________________________________________
TheoryLunch mailing list
TheoryLunch at weizmann.ac.il
https://mailman.weizmann.ac.il/mailman/listinfo/theorylunch
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
From gvaliant at cs.stanford.edu Wed Apr 3 10:19:02 2019
From: gvaliant at cs.stanford.edu (Gregory Valiant)
Date: Wed, 3 Apr 2019 10:19:02 -0700
Subject: [theory-seminar] Fake FOCS submissions site!!!! [fwd]
Message-ID:
Just a heads up for those of you submitting papers to FOCS....the first or
second google hit is a fake submission server.
Begin forwarded message:
????????????
As April 5 (FOCS submission deadline) approaches:
Beware of a site (aconf dot org) that pretends to be FOCS and runs a fake
submission server!
When googling "FOCS 2019", it shows up first or second (lower on bing)
Purpose unknown, phishing passwords?
????????????
I don?t know for sure that the intention is malicious, and I and others
have taken other steps regarding the potential threat, but it?s a good idea
to have as many people aware of this as possible. We (FOCS 2019 team) are
aware of at least two cases where people fell for this and noticed it. We
don?t know how many people fell for this and did not notice it.
Best and thanks for your help,
Yuval
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
From shivamgarg at stanford.edu Wed Apr 3 14:17:32 2019
From: shivamgarg at stanford.edu (Shivam Garg)
Date: Wed, 3 Apr 2019 21:17:32 +0000
Subject: [theory-seminar] Algorithms and Friends Lunch,
Monday (April 8) at Gates 463A
Message-ID:
Hi everyone,
We have an exciting lineup of speakers for algorithms and friends lunch this quarter, with Ravi Kumar (Google) speaking on coming Monday (April 8), at noon, in Gates 463A. Food will be there! I will send the abstract by the end of this week.
The list of confirmed speakers can be found here:
http://theory.stanford.edu/algofriends/possible_dates.html
Thanks,
Shivam
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
From ofirgeri at stanford.edu Thu Apr 4 11:17:51 2019
From: ofirgeri at stanford.edu (Ofir Geri)
Date: Thu, 4 Apr 2019 18:17:51 +0000
Subject: [theory-seminar] Course Announcement: Graduate Course on Mechanism
Design
Message-ID:
Hi all,
Itai Ashlagi will be teaching a graduate course on mechanism and market design starting April 9. The syllabus is attached.
Best,
Ofir
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
-------------- next part --------------
A non-text attachment was scrubbed...
Name: MSE-365-syllabus.pdf
Type: application/pdf
Size: 60680 bytes
Desc: MSE-365-syllabus.pdf
URL:
From ccanonne at cs.stanford.edu Thu Apr 4 14:04:32 2019
From: ccanonne at cs.stanford.edu (=?UTF-8?Q?Cl=c3=a9ment_Canonne?=)
Date: Thu, 4 Apr 2019 14:04:32 -0700
Subject: [theory-seminar] Lecture in Simons: Randomness in Number Theory
Message-ID:
Hi everyone,
Just FYI, there is this (general-audience) event as Simons, on April
15th, we looks rather interesting and might be of interest:
https://simons.berkeley.edu/events/openlectures2019-spring-3
> Randomness in Number Theory, by Peter Sarnak (Princeton University)
> By way of concrete examples, we discuss the dichotomy that in number
> theory many basic phenomena are either very structured and if not,
> then they are random. The models for randomness for different
> problems can be quite unexpected, and understanding and establishing
> the randomness is often the key issue. Conversely the fact that
> certain number-theoretic quantities behave randomly is a powerful
> source for the construction of much sought-after explicit
> pseudo-random objects.
Best,
--
-- Cl?ment
From ofirgeri at stanford.edu Fri Apr 5 11:39:58 2019
From: ofirgeri at stanford.edu (Ofir Geri)
Date: Fri, 5 Apr 2019 18:39:58 +0000
Subject: [theory-seminar] Theory Seminar (4/5): Shant Boodaghians
In-Reply-To:
References:
Message-ID:
Reminder: Shant's talk is today at 3pm.
________________________________
From: Ofir Geri
Sent: Monday, April 1, 2019 3:56:47 PM
To: thseminar at cs.stanford.edu
Subject: Theory Seminar (4/5): Shant Boodaghians
Hi all,
The first theory seminar of the quarter will be given by Shant Boodaghians (UIUC) on Reductions and Algorithms for Pure Nash Equilibria in Smoothed Network Coordination Games (see abstract below). The talk will be as usual on Friday, 3:00 PM in Gates 463A.
Hope to see you there!
Ofir
Reductions and Algorithms for Pure Nash Equilibria in Smoothed Network Coordination Games
Speaker: Shant Boodaghians (UIUC)
A 2-player coordination game is a game where both players receive the same payoff, and are therefore collaborating. A network coordination game is a multi-player game where the players are nodes in a graph, and each edge models a 2-player coordination game. The strategy chosen by each node applies simultaneously for all incident edges. We study the PLS-complete problem of finding a pure-strategy Nash equilibrium in such games, when the instance is smoothed. (The payoff values are independent random variables, where the distribution is chosen adversarially under a density constraint.) We show that the natural best-response dynamic finds an equilibrium in polynomial time when the number of strategies that each player may choose from is a constant.
We first introduce smoothness-preserving reductions, and show that the problem admits such a reduction to finding a locally maximal cut in a weigthed graph, which allows us to inherit recent developments in the smoothed analysis of the local-search algorithm for this problem. Note that local-max-cut is also PLS-complete. This reduction, however, only holds when each player is allowed to choose from exactly 2 strategies. In the general case, we show a reduction to finding a cut maximal up to flipping 2 nodes; this variant has not been shown to be efficient in the smoothed case, and so the reduction is conditional.
Second, we explain how the framework for the analysis of local-max-cut may be applied directly to our problem, rather than going through reductions. When players are allowed to choose from k strategies, we show that best-response searches converge in polynomial time, where the exponent is linear in k. The conditional reduction above seeks to eliminate the exponential dependence on k.
(Joint work with Rucha Kulkarni and Ruta Mehta)
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
From gvaliant at cs.stanford.edu Fri Apr 5 13:22:54 2019
From: gvaliant at cs.stanford.edu (Gregory Valiant)
Date: Fri, 5 Apr 2019 13:22:54 -0700
Subject: [theory-seminar] post-deadline happy hour!
Message-ID:
Hi Friends,
As usual, we will be having a post FOCS-deadline happy hour today, @4pm
(immediately after the seminar). Note the deadline is actually 3pm!
Unusually, I will be missing this happy hour (the first post FOCS/STOC
deadline happy hour that I've missed since I started at Stanford...this
includes the 10am deadline a few years ago when I was the only one
celebrating), as I have a flight @5pm. That said, I will be drinking a
beer with you all in spirit. I dropped off some beer in Mary's office,
which will find its way to the theory fridge....
Enjoy!
-g
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
From shivamgarg at stanford.edu Sun Apr 7 09:52:28 2019
From: shivamgarg at stanford.edu (Shivam Garg)
Date: Sun, 7 Apr 2019 16:52:28 +0000
Subject: [theory-seminar] Algorithms and Friends Lunch,
Monday (April 8) at Gates 463A
In-Reply-To:
References:
Message-ID:
Hi everyone,
Ravi Kumar (Google) will be giving a talk tomorrow (April 8), at noon, in Gates 463A.
Title: Random Walks and Network Properties
Abstract: A random walk is a natural way to explore a network. We will study the use of uniform random walks to estimate various properties such as the size of the network, average degree, number of triangles, etc. Less obvious random walks can also be designed to do other tasks such as uniformly generating a node or counting network motifs. However, our perspective is that one has to be careful in using random walks for applications.
Bio: Ravi Kumar has been a senior staff research scientist at Google since June 2012. Prior to this, he was a research staff member at the IBM Almaden Research Center and a principal research scientist at Yahoo! Research. He obtained his PhD in Computer Science from Cornell University in 1998. His research interests include Web search and data mining, algorithms for massive data, and the theory of computation.
--Shivam
________________________________
From: Shivam Garg
Sent: Thursday, April 4, 2019 2:47 AM
To: algorithms-and-friends at lists.stanford.edu; theory-seminar at lists.stanford.edu
Subject: Algorithms and Friends Lunch, Monday (April 8) at Gates 463A
Hi everyone,
We have an exciting lineup of speakers for algorithms and friends lunch this quarter, with Ravi Kumar (Google) speaking on coming Monday (April 8), at noon, in Gates 463A. Food will be there! I will send the abstract by the end of this week.
The list of confirmed speakers can be found here:
http://theory.stanford.edu/algofriends/possible_dates.html
Thanks,
Shivam
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
From shivamgarg at stanford.edu Mon Apr 8 11:27:17 2019
From: shivamgarg at stanford.edu (Shivam Garg)
Date: Mon, 8 Apr 2019 18:27:17 +0000
Subject: [theory-seminar] Algorithms and Friends Lunch,
Monday (April 8) at Gates 463A
In-Reply-To:
References: ,
Message-ID:
Reminder, this is starting in 30 mins!
________________________________
From: Shivam Garg
Sent: Sunday, April 7, 2019 9:52:28 AM
To: algorithms-and-friends at lists.stanford.edu; theory-seminar at lists.stanford.edu
Cc: Megan D. Harris
Subject: Re: Algorithms and Friends Lunch, Monday (April 8) at Gates 463A
Hi everyone,
Ravi Kumar (Google) will be giving a talk tomorrow (April 8), at noon, in Gates 463A.
Title: Random Walks and Network Properties
Abstract: A random walk is a natural way to explore a network. We will study the use of uniform random walks to estimate various properties such as the size of the network, average degree, number of triangles, etc. Less obvious random walks can also be designed to do other tasks such as uniformly generating a node or counting network motifs. However, our perspective is that one has to be careful in using random walks for applications.
Bio: Ravi Kumar has been a senior staff research scientist at Google since June 2012. Prior to this, he was a research staff member at the IBM Almaden Research Center and a principal research scientist at Yahoo! Research. He obtained his PhD in Computer Science from Cornell University in 1998. His research interests include Web search and data mining, algorithms for massive data, and the theory of computation.
--Shivam
________________________________
From: Shivam Garg
Sent: Thursday, April 4, 2019 2:47 AM
To: algorithms-and-friends at lists.stanford.edu; theory-seminar at lists.stanford.edu
Subject: Algorithms and Friends Lunch, Monday (April 8) at Gates 463A
Hi everyone,
We have an exciting lineup of speakers for algorithms and friends lunch this quarter, with Ravi Kumar (Google) speaking on coming Monday (April 8), at noon, in Gates 463A. Food will be there! I will send the abstract by the end of this week.
The list of confirmed speakers can be found here:
http://theory.stanford.edu/algofriends/possible_dates.html
Thanks,
Shivam
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
From ofirgeri at stanford.edu Mon Apr 8 21:25:00 2019
From: ofirgeri at stanford.edu (Ofir Geri)
Date: Tue, 9 Apr 2019 04:25:00 +0000
Subject: [theory-seminar] Theory Seminar (4/12): Alexandra Kolla
Message-ID:
Hi all,
This week's theory seminar talk will be given by Alexandra Kolla from CU Boulder (see abstract below). The talk will be as usual on Friday, 3:00pm in Gates 463A.
Hope to see you there!
Ofir
A Ramsey-type Theorem on the Max-Cut Value of d-Regular Graphs
Speaker: Alexandra Kolla (CU Boulder)
We consider the problem of finding large cuts in graphs that contain no small cliques. We show that for any r >= 3, every k_r-free graphs with maximum degree d has a cut that cuts at least a 1/2 + ?(1/d^{1-1/2^{r-2}}) fraction of its edges. This generalizes a result of Shearer that shows every triangle-free d-regular graph has a cut that cuts at least a 1/2 + ?(1/\sqrt{d}) fraction of its edges. Our proof yields a randomized polynomial-time algorithm.
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
From wyma at stanford.edu Tue Apr 9 12:04:00 2019
From: wyma at stanford.edu (Weiyun Ma)
Date: Tue, 9 Apr 2019 19:04:00 +0000
Subject: [theory-seminar] Theory Lunch 4/11 -- Don Knuth
Message-ID:
Hi All,
This Thursday at theory lunch, Don Knuth will tell us about "Ball-piling Puzzles." As usual, we meet from noon to 1pm at 463A.
If you are interested in giving a talk sometime this quarter, you can sign up at:
https://docs.google.com/document/d/1S0QcDMTn-JRaP1cRFRihZyeNfHUmpYLkyORKMcGIBgY/edit
Best,
Anna
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
From wyma at stanford.edu Tue Apr 9 12:04:00 2019
From: wyma at stanford.edu (Weiyun Ma)
Date: Tue, 9 Apr 2019 19:04:00 +0000
Subject: [theory-seminar] Theory Lunch 4/11 -- Don Knuth
Message-ID:
Hi All,
This Thursday at theory lunch, Don Knuth will tell us about "Ball-piling Puzzles." As usual, we meet from noon to 1pm at 463A.
If you are interested in giving a talk sometime this quarter, you can sign up at:
https://docs.google.com/document/d/1S0QcDMTn-JRaP1cRFRihZyeNfHUmpYLkyORKMcGIBgY/edit
Best,
Anna
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
From shivamgarg at stanford.edu Thu Apr 11 08:43:12 2019
From: shivamgarg at stanford.edu (Shivam Garg)
Date: Thu, 11 Apr 2019 15:43:12 +0000
Subject: [theory-seminar] Algorithms and Friends Lunch,
Monday (April 15) at Gates 463A
Message-ID:
Hi everyone,
We are excited to have Deborah M Gordon speak at the Algorithms and Friends lunch, on coming Monday (April 15), at noon, in Gates 463A.
Title: The evolution of the algorithms for collective behavior.
Bio: Deborah M. Gordon is a Professor in the Department of Biology at Stanford University. She studies how ant colonies work without central control using networks of simple interactions, and how these networks evolve in relation to changing environments. She received her PhD from Duke University, then joined the Harvard Society of Fellows, and did postdoctoral research at Oxford and the University of London before joining the Stanford faculty in 1991. Projects include a long-term study of a population of harvester ant colonies in Arizona, studies of the invasive Argentine ant in northern California, arboreal ant trail networks and ant-plant mutualisms in Central America.
Thanks,
Shivam
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
From ofirgeri at stanford.edu Fri Apr 12 11:49:42 2019
From: ofirgeri at stanford.edu (Ofir Geri)
Date: Fri, 12 Apr 2019 18:49:42 +0000
Subject: [theory-seminar] Theory Seminar (4/12): Alexandra Kolla
In-Reply-To:
References:
Message-ID:
Reminder: Alexandra's talk is today at 3pm.
________________________________
From: Ofir Geri
Sent: Monday, April 8, 2019 9:25 PM
To: thseminar at cs.stanford.edu
Subject: Theory Seminar (4/12): Alexandra Kolla
Hi all,
This week's theory seminar talk will be given by Alexandra Kolla from CU Boulder (see abstract below). The talk will be as usual on Friday, 3:00pm in Gates 463A.
Hope to see you there!
Ofir
A Ramsey-type Theorem on the Max-Cut Value of d-Regular Graphs
Speaker: Alexandra Kolla (CU Boulder)
We consider the problem of finding large cuts in graphs that contain no small cliques. We show that for any r >= 3, every k_r-free graphs with maximum degree d has a cut that cuts at least a 1/2 + ?(1/d^{1-1/2^{r-2}}) fraction of its edges. This generalizes a result of Shearer that shows every triangle-free d-regular graph has a cut that cuts at least a 1/2 + ?(1/\sqrt{d}) fraction of its edges. Our proof yields a randomized polynomial-time algorithm.
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
From ccanonne at stanford.edu Sat Apr 13 07:20:00 2019
From: ccanonne at stanford.edu (Clement Louis Arthur Canonne)
Date: Sat, 13 Apr 2019 14:20:00 +0000
Subject: [theory-seminar] Fwd: 4/17 A Woman's Guide to Negotiating - with
Susie Brubaker-Cole [event][misc]
In-Reply-To: <68DD1C20-D290-4C6E-BDC1-922F50D40D5E@stanford.edu>
References: <8D98080C-452A-42B5-99A9-26F3C7BC6C8D@stanford.edu>,
,
<68DD1C20-D290-4C6E-BDC1-922F50D40D5E@stanford.edu>
Message-ID:
FYI (it could be of interest).
-- Cl?ment
---------- Forwarded message ----------
From: Ioana Alexandra Marin
Date: Apr 12, 2019 8:17 PM
Subject: 4/17 A Woman's Guide to Negotiating - with Susie Brubaker-Cole [event][misc]
To: postdoc-exchange at lists.stanford.edu
Cc:
Howdy, everyone! A week from today we are presenting a very special event at the Women?s Community Center.
Please help us spread the word by directing folks to our events page, forwarding the blurb below, and/or disseminating the attached flyer. THANK YOU!
A Woman's Guide to Negotiating: A Conversation with Susie Brubaker-Cole
Wednesday, April 17, 2019 5:15pm-6:30pm at the Women?s Community Center
Join us for this unique opportunity have a conversation with Vice Provost for Student Affairs, Susie Brubaker-Cole. The topic is negotiating. She is excited for this chance to share her super-top-secret-fool-proof negotiation tips to add to your toolkit of skills for your financial wellness journey. Don?t miss it.
Please RSVP for Dinner
Angela Amarillas, ?97, MA ?01 | Program Manager, Financial Literacy
Stanford Student Financial Services | 408 Panama Mall, Suite 217
415.846.7368 | mindovermoney.stanford.edu
_______________________________________________
mindovermoney_financial_literacy mailing list
mindovermoney_financial_literacy at lists.stanford.edu
https://mailman.stanford.edu/mailman/listinfo/mindovermoney_financial_literacy
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
-------------- next part --------------
A non-text attachment was scrubbed...
Name: 4-17-19 WCC Susie Brubaker-Cole.pdf
Type: application/pdf
Size: 203890 bytes
Desc: 4-17-19 WCC Susie Brubaker-Cole.pdf
URL:
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
-------------- next part --------------
An embedded and charset-unspecified text was scrubbed...
Name: ATT00001.txt
URL:
From ofirgeri at stanford.edu Sat Apr 13 12:49:39 2019
From: ofirgeri at stanford.edu (Ofir Geri)
Date: Sat, 13 Apr 2019 19:49:39 +0000
Subject: [theory-seminar] Theory Seminar (4/19): Michel Schellekens
Message-ID:
Hi all,
The next theory seminar will be on Friday (4/19). Michel Schellekens (University College Cork) will give a talk on Thermodynamic algorithms: Duality and Entropy Conservation (see abstract below). The talk will be as usual at 3:00pm in Gates 463A.
Hope to see you there!
Ofir
Thermodynamic algorithms: Duality and Entropy Conservation
Speaker: Michel Schellekens (University College Cork)
We introduce a new type of algebra, the Expedient Algebra EXP, for which computations satisfy tight distribution control. Algorithms satisfying such distribution control are guaranteed to support modular time analysis?drastically simplifying static timing. The property ensures that problematic algorithms, for which the exact time is too hard or impossible to analyze with current means, can be distinguished at code level from algorithms supporting an elegant modular time analysis. Little is known about the intrinsic properties of algorithms exhibiting such distribution control. We show that EXP-computations, made reversible through history-keeping, act as closed systems in which entropy is conserved. We refer to such algorithms as "thermodynamic algorithms". Thus modularity of timing is linked to entropy conservation of data flow, sharpening traditional entropy preservation guaranteed by the second law of thermodynamics for reversible systems. Conservation typically holds for the case of energy, but not for entropy. A salient point is that for algorithms satisfying distribution control, entropy is neither created nor destroyed, merely transferred from one form, i.e. quantitative entropy, to another, i.e. positional entropy.
We establish an Entropy Conservation Law ECL. The law expresses the inverse proportionality of positional and quantitative entropy for distributions over series-parallel orders and their duals. We also present a duality theorem which expresses that order established by the expedient product (with history) on labels is proportionally destroyed on indices by a shadow transformation in the dual space. The duality theorem supports the derivation of a computational version of ECL (obtained jointly with M. Fiore during a recent research visit at Cambridge).
The results shed new light on the properties of algorithms for which distribution control, and hence modular timing, is guaranteed.
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
From mpcs at neptune.ucc.ie Sun Apr 14 12:54:14 2019
From: mpcs at neptune.ucc.ie (mpcs at neptune.ucc.ie)
Date: Sun, 14 Apr 2019 19:54:14 +0000
Subject: [theory-seminar] theory-seminar Digest, Vol 56, Issue 14
In-Reply-To:
References:
Message-ID: <22719dda91584b6c9b0db2cef6df783f@nocturne.ucc.ie>
Thanks Ofir,
I shared a LinkedIn article on the talk and its topic (for a broader audience). It includes some further info and intuitions on the subject.
https://www.linkedin.com/pulse/what-messy-offices-have-teach-us-algorithm-design-michel-schellekens/
Best wishes,
Michel
April 14, 2019 12:24 PM, theory-seminar-request at lists.stanford.edu wrote:
> Send theory-seminar mailing list submissions to
> theory-seminar at lists.stanford.edu
>
> To subscribe or unsubscribe via the World Wide Web, visit
> https://mailman.stanford.edu/mailman/listinfo/theory-seminar
> or, via email, send a message with subject or body 'help' to
> theory-seminar-request at lists.stanford.edu
>
> You can reach the person managing the list at
> theory-seminar-owner at lists.stanford.edu
>
> When replying, please edit your Subject line so it is more specific
> than "Re: Contents of theory-seminar digest..."
>
> Today's Topics:
>
> 1. Theory Seminar (4/19): Michel Schellekens (Ofir Geri)
>
> ----------------------------------------------------------------------
>
> Message: 1
> Date: Sat, 13 Apr 2019 19:49:39 +0000
> From: Ofir Geri
> To: "thseminar at cs.stanford.edu"
> Subject: [theory-seminar] Theory Seminar (4/19): Michel Schellekens
> Message-ID:
>
>
> Content-Type: text/plain; charset="windows-1252"
>
> Hi all,
>
> The next theory seminar will be on Friday (4/19). Michel Schellekens (University College Cork) will
> give a talk on Thermodynamic algorithms: Duality and Entropy Conservation (see abstract below). The
> talk will be as usual at 3:00pm in Gates 463A.
>
> Hope to see you there!
> Ofir
>
> Thermodynamic algorithms: Duality and Entropy Conservation
> Speaker: Michel Schellekens (University College Cork)
>
> We introduce a new type of algebra, the Expedient Algebra EXP, for which computations satisfy tight
> distribution control. Algorithms satisfying such distribution control are guaranteed to support
> modular time analysis?drastically simplifying static timing. The property ensures that problematic
> algorithms, for which the exact time is too hard or impossible to analyze with current means, can
> be distinguished at code level from algorithms supporting an elegant modular time analysis. Little
> is known about the intrinsic properties of algorithms exhibiting such distribution control. We show
> that EXP-computations, made reversible through history-keeping, act as closed systems in which
> entropy is conserved. We refer to such algorithms as "thermodynamic algorithms". Thus modularity of
> timing is linked to entropy conservation of data flow, sharpening traditional entropy preservation
> guaranteed by the second law of thermodynamics for reversible systems. Conservation typically holds
> for the case
> of energy, but not for entropy. A salient point is that for algorithms satisfying distribution
> control, entropy is neither created nor destroyed, merely transferred from one form, i.e.
> quantitative entropy, to another, i.e. positional entropy.
>
> We establish an Entropy Conservation Law ECL. The law expresses the inverse proportionality of
> positional and quantitative entropy for distributions over series-parallel orders and their duals.
> We also present a duality theorem which expresses that order established by the expedient product
> (with history) on labels is proportionally destroyed on indices by a shadow transformation in the
> dual space. The duality theorem supports the derivation of a computational version of ECL (obtained
> jointly with M. Fiore during a recent research visit at Cambridge).
>
> The results shed new light on the properties of algorithms for which distribution control, and
> hence modular timing, is guaranteed.
> -------------- next part --------------
> An HTML attachment was scrubbed...
> URL:
> tml>
>
> ------------------------------
>
> Subject: Digest Footer
>
> _______________________________________________
> theory-seminar mailing list
> theory-seminar at lists.stanford.edu
> https://mailman.stanford.edu/mailman/listinfo/theory-seminar
>
> ------------------------------
>
> End of theory-seminar Digest, Vol 56, Issue 14
> **********************************************
From ofirgeri at stanford.edu Sun Apr 14 22:32:55 2019
From: ofirgeri at stanford.edu (Ofir Geri)
Date: Mon, 15 Apr 2019 05:32:55 +0000
Subject: [theory-seminar] Theory Seminar (4/19): Michel Schellekens
In-Reply-To:
References:
Message-ID:
If you'd like get more information about the upcoming talk, the following article by Michel includes further details and intuitions on the subject (for a broader audience):
https://www.linkedin.com/pulse/what-messy-offices-have-teach-us-algorithm-design-michel-schellekens/
________________________________
From: Ofir Geri
Sent: Saturday, April 13, 2019 12:49 PM
To: thseminar at cs.stanford.edu
Subject: Theory Seminar (4/19): Michel Schellekens
Hi all,
The next theory seminar will be on Friday (4/19). Michel Schellekens (University College Cork) will give a talk on Thermodynamic algorithms: Duality and Entropy Conservation (see abstract below). The talk will be as usual at 3:00pm in Gates 463A.
Hope to see you there!
Ofir
Thermodynamic algorithms: Duality and Entropy Conservation
Speaker: Michel Schellekens (University College Cork)
We introduce a new type of algebra, the Expedient Algebra EXP, for which computations satisfy tight distribution control. Algorithms satisfying such distribution control are guaranteed to support modular time analysis?drastically simplifying static timing. The property ensures that problematic algorithms, for which the exact time is too hard or impossible to analyze with current means, can be distinguished at code level from algorithms supporting an elegant modular time analysis. Little is known about the intrinsic properties of algorithms exhibiting such distribution control. We show that EXP-computations, made reversible through history-keeping, act as closed systems in which entropy is conserved. We refer to such algorithms as "thermodynamic algorithms". Thus modularity of timing is linked to entropy conservation of data flow, sharpening traditional entropy preservation guaranteed by the second law of thermodynamics for reversible systems. Conservation typically holds for the case of energy, but not for entropy. A salient point is that for algorithms satisfying distribution control, entropy is neither created nor destroyed, merely transferred from one form, i.e. quantitative entropy, to another, i.e. positional entropy.
We establish an Entropy Conservation Law ECL. The law expresses the inverse proportionality of positional and quantitative entropy for distributions over series-parallel orders and their duals. We also present a duality theorem which expresses that order established by the expedient product (with history) on labels is proportionally destroyed on indices by a shadow transformation in the dual space. The duality theorem supports the derivation of a computational version of ECL (obtained jointly with M. Fiore during a recent research visit at Cambridge).
The results shed new light on the properties of algorithms for which distribution control, and hence modular timing, is guaranteed.
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
From wyma at stanford.edu Sun Apr 14 21:32:22 2019
From: wyma at stanford.edu (Weiyun Ma)
Date: Mon, 15 Apr 2019 04:32:22 +0000
Subject: [theory-seminar] Theory Lunch 4/18?
Message-ID:
Hi everyone,
Just a note that the slot for the theory lunch this coming Thursday (4/18) is still open. If you cannot wait to share an intriguing problem or an interesting idea, or if you want to organize fun activities with theory peeps, please let me know, and the stage is yours!
Best,
Anna
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
From moses at cs.stanford.edu Mon Apr 15 01:03:49 2019
From: moses at cs.stanford.edu (Moses Charikar)
Date: Mon, 15 Apr 2019 01:03:49 -0700
Subject: [theory-seminar] Jan Vondrak at Probability Seminar today
Message-ID:
Folks,
Jan Vondrak is speaking today (Monday) at the Probability Seminar on
"A concentration inequality for generalization error of uniformly stable
algorithms".
The abstract is here:
https://statistics.stanford.edu/sites/g/files/sbiybj6031/f/Apr15-2019.pdf
It should be of interest to many of you on this list.
This is at 4pm, Monday, April 15, 2019 Sequoia Hall Room 200
Best,
Moses
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
From shivamgarg at stanford.edu Mon Apr 15 11:28:04 2019
From: shivamgarg at stanford.edu (Shivam Garg)
Date: Mon, 15 Apr 2019 18:28:04 +0000
Subject: [theory-seminar] Algorithms and Friends Lunch,
Monday (April 15) at Gates 463A
In-Reply-To:
References:
Message-ID:
Reminder, this is starting in 30 mins!
________________________________
From: Shivam Garg
Sent: Thursday, April 11, 2019 8:43:12 AM
To: algorithms-and-friends at lists.stanford.edu; theory-seminar at lists.stanford.edu
Cc: Megan D. Harris; Deborah M Gordon
Subject: Algorithms and Friends Lunch, Monday (April 15) at Gates 463A
Hi everyone,
We are excited to have Deborah M Gordon speak at the Algorithms and Friends lunch, on coming Monday (April 15), at noon, in Gates 463A.
Title: The evolution of the algorithms for collective behavior.
Bio: Deborah M. Gordon is a Professor in the Department of Biology at Stanford University. She studies how ant colonies work without central control using networks of simple interactions, and how these networks evolve in relation to changing environments. She received her PhD from Duke University, then joined the Harvard Society of Fellows, and did postdoctoral research at Oxford and the University of London before joining the Stanford faculty in 1991. Projects include a long-term study of a population of harvester ant colonies in Arizona, studies of the invasive Argentine ant in northern California, arboreal ant trail networks and ant-plant mutualisms in Central America.
Thanks,
Shivam
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
From wyma at stanford.edu Wed Apr 17 08:00:02 2019
From: wyma at stanford.edu (Weiyun Ma)
Date: Wed, 17 Apr 2019 15:00:02 +0000
Subject: [theory-seminar] Theory Lunch 4/18?
In-Reply-To:
References:
Message-ID:
To follow up, the slot for theory lunch tomorrow is still available for grabs! Let me know if you would like to take it.
Anna
________________________________
From: theory-seminar on behalf of Weiyun Ma
Sent: Sunday, April 14, 2019 9:32:22 PM
To: thseminar at cs.stanford.edu
Subject: [theory-seminar] Theory Lunch 4/18?
Hi everyone,
Just a note that the slot for the theory lunch this coming Thursday (4/18) is still open. If you cannot wait to share an intriguing problem or an interesting idea, or if you want to organize fun activities with theory peeps, please let me know, and the stage is yours!
Best,
Anna
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
From wyma at stanford.edu Wed Apr 17 12:17:06 2019
From: wyma at stanford.edu (Weiyun Ma)
Date: Wed, 17 Apr 2019 19:17:06 +0000
Subject: [theory-seminar] Theory Lunch 4/18?
In-Reply-To:
References: ,
Message-ID:
Update: Weihao has kindly offered to give a talk tomorrow! As always, we meet from noon to 1pm at 463A.
----------------------------------------------------------
Maximum Likelihood Estimation for Learning Populations of Parameters
Speaker: Weihao Kong
Consider the following estimation problem: there are N entities, each with an unknown parameter p_i in [0,1], drawn from an unknown distribution P, and we observe N independent random variables, X_1,...,X_N, with X_i~Binomial(t, p_i). How accurately can one recover P? This problem arises in numerous domains, including "federated learning" settings, and biological and medical settings, where the size of the population under study, N, is large in comparison to the number of observations per individual, t. In this work, we show that the maximum likelihood estimator is both statistically minimax optimal and efficiently computable. Precisely, the MLE achieves the information theoretic optimal error bound of max(1/t,1/sqrt(t logN)), with regards to the Wasserstein distance. This improves on the performance of the previous 'method of moments' approach when N < exp(t), and is significantly better than the 1/sqrt(t) error of the naive empirical estimator.
----------------------------------------------------------
Best,
Anna
________________________________
From: theory-seminar on behalf of Weiyun Ma
Sent: Wednesday, April 17, 2019 8:00:02 AM
To: thseminar at cs.stanford.edu
Subject: Re: [theory-seminar] Theory Lunch 4/18?
To follow up, the slot for theory lunch tomorrow is still available for grabs! Let me know if you would like to take it.
Anna
________________________________
From: theory-seminar on behalf of Weiyun Ma
Sent: Sunday, April 14, 2019 9:32:22 PM
To: thseminar at cs.stanford.edu
Subject: [theory-seminar] Theory Lunch 4/18?
Hi everyone,
Just a note that the slot for the theory lunch this coming Thursday (4/18) is still open. If you cannot wait to share an intriguing problem or an interesting idea, or if you want to organize fun activities with theory peeps, please let me know, and the stage is yours!
Best,
Anna
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
From ofirgeri at stanford.edu Fri Apr 19 09:35:05 2019
From: ofirgeri at stanford.edu (Ofir Geri)
Date: Fri, 19 Apr 2019 16:35:05 +0000
Subject: [theory-seminar] Theory Seminar (4/19): Michel Schellekens
In-Reply-To:
References: ,
Message-ID:
Reminder: Michel's talk is today at 3pm.
________________________________
From: Ofir Geri
Sent: Sunday, April 14, 2019 10:32 PM
To: thseminar at cs.stanford.edu
Subject: Re: Theory Seminar (4/19): Michel Schellekens
If you'd like get more information about the upcoming talk, the following article by Michel includes further details and intuitions on the subject (for a broader audience):
https://www.linkedin.com/pulse/what-messy-offices-have-teach-us-algorithm-design-michel-schellekens/
________________________________
From: Ofir Geri
Sent: Saturday, April 13, 2019 12:49 PM
To: thseminar at cs.stanford.edu
Subject: Theory Seminar (4/19): Michel Schellekens
Hi all,
The next theory seminar will be on Friday (4/19). Michel Schellekens (University College Cork) will give a talk on Thermodynamic algorithms: Duality and Entropy Conservation (see abstract below). The talk will be as usual at 3:00pm in Gates 463A.
Hope to see you there!
Ofir
Thermodynamic algorithms: Duality and Entropy Conservation
Speaker: Michel Schellekens (University College Cork)
We introduce a new type of algebra, the Expedient Algebra EXP, for which computations satisfy tight distribution control. Algorithms satisfying such distribution control are guaranteed to support modular time analysis?drastically simplifying static timing. The property ensures that problematic algorithms, for which the exact time is too hard or impossible to analyze with current means, can be distinguished at code level from algorithms supporting an elegant modular time analysis. Little is known about the intrinsic properties of algorithms exhibiting such distribution control. We show that EXP-computations, made reversible through history-keeping, act as closed systems in which entropy is conserved. We refer to such algorithms as "thermodynamic algorithms". Thus modularity of timing is linked to entropy conservation of data flow, sharpening traditional entropy preservation guaranteed by the second law of thermodynamics for reversible systems. Conservation typically holds for the case of energy, but not for entropy. A salient point is that for algorithms satisfying distribution control, entropy is neither created nor destroyed, merely transferred from one form, i.e. quantitative entropy, to another, i.e. positional entropy.
We establish an Entropy Conservation Law ECL. The law expresses the inverse proportionality of positional and quantitative entropy for distributions over series-parallel orders and their duals. We also present a duality theorem which expresses that order established by the expedient product (with history) on labels is proportionally destroyed on indices by a shadow transformation in the dual space. The duality theorem supports the derivation of a computational version of ECL (obtained jointly with M. Fiore during a recent research visit at Cambridge).
The results shed new light on the properties of algorithms for which distribution control, and hence modular timing, is guaranteed.
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
From jbrakens at stanford.edu Fri Apr 19 09:59:37 2019
From: jbrakens at stanford.edu (Joshua Brakensiek)
Date: Fri, 19 Apr 2019 16:59:37 +0000
Subject: [theory-seminar] Call for Stanford theory website updates and
suggestions
Message-ID: <145F4FEE-5678-4B4B-868B-78B18C76DA24@stanford.edu>
Hi all,
I am a first year PhD student in the theory group. I?ve recently taken over the role of webmaster of theory.stanford.edu . I am actively looking for suggestions as to how I can best improve the theory website. In particular,
1) If there is anything on the website about you (or someone you know) that is incorrect, please let me know.
2) If there are other ongoing issues about the website that you are aware of (e.g., certain parts are disorganized, the website looks out of date, etc.), please let me know so that I can address them.
3) I?m also open to any longer-term suggestions on how the website could be more beneficial to people inside and outside of Stanford.
Thank you,
Josh
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
From marykw at stanford.edu Fri Apr 19 14:59:08 2019
From: marykw at stanford.edu (Mary Wootters)
Date: Fri, 19 Apr 2019 14:59:08 -0700
Subject: [theory-seminar] Theory Seminar (4/19): Michel Schellekens
In-Reply-To:
References:
Message-ID:
Another reminder that this is in 3 minutes, along with an advertisement
that Michel has brought some home-made dessert! :)
--Mary
On Fri, Apr 19, 2019 at 9:35 AM Ofir Geri wrote:
> Reminder: Michel's talk is today at 3pm.
> ------------------------------
> *From:* Ofir Geri
> *Sent:* Sunday, April 14, 2019 10:32 PM
> *To:* thseminar at cs.stanford.edu
> *Subject:* Re: Theory Seminar (4/19): Michel Schellekens
>
> If you'd like get more information about the upcoming talk, the following
> article by Michel includes further details and intuitions on the subject (for
> a broader audience):
>
> https://www.linkedin.com/pulse/what-messy-offices-have-teach-us-algorithm-design-michel-schellekens/
> ------------------------------
> *From:* Ofir Geri
> *Sent:* Saturday, April 13, 2019 12:49 PM
> *To:* thseminar at cs.stanford.edu
> *Subject:* Theory Seminar (4/19): Michel Schellekens
>
> Hi all,
>
> The next theory seminar will be on Friday (4/19). Michel Schellekens
> (University College Cork) will give a talk on Thermodynamic algorithms:
> Duality and Entropy Conservation (see abstract below). The talk will be as
> usual at 3:00pm in Gates 463A.
>
> Hope to see you there!
> Ofir
>
> *Thermodynamic algorithms: Duality and Entropy Conservation*
> Speaker: Michel Schellekens (University College Cork)
>
> We introduce a new type of algebra, the Expedient Algebra EXP, for which
> computations satisfy tight distribution control. Algorithms satisfying such
> distribution control are guaranteed to support modular time
> analysis?drastically simplifying static timing. The property ensures that
> problematic algorithms, for which the exact time is too hard or impossible
> to analyze with current means, can be distinguished at code level from
> algorithms supporting an elegant modular time analysis. Little is known
> about the intrinsic properties of algorithms exhibiting such distribution
> control. We show that EXP-computations, made reversible through
> history-keeping, act as closed systems in which entropy is conserved. We
> refer to such algorithms as "thermodynamic algorithms". Thus modularity of
> timing is linked to entropy conservation of data flow, sharpening
> traditional entropy preservation guaranteed by the second law of
> thermodynamics for reversible systems. Conservation typically holds for the
> case of energy, but not for entropy. A salient point is that for algorithms
> satisfying distribution control, entropy is neither created nor destroyed,
> merely transferred from one form, i.e. quantitative entropy, to another,
> i.e. positional entropy.
>
> We establish an Entropy Conservation Law ECL. The law expresses the
> inverse proportionality of positional and quantitative entropy for
> distributions over series-parallel orders and their duals. We also present
> a duality theorem which expresses that order established by the expedient
> product (with history) on labels is proportionally destroyed on indices by
> a shadow transformation in the dual space. The duality theorem supports the
> derivation of a computational version of ECL (obtained jointly with M.
> Fiore during a recent research visit at Cambridge).
>
> The results shed new light on the properties of algorithms for which
> distribution control, and hence modular timing, is guaranteed.
> _______________________________________________
> 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 Apr 22 22:48:42 2019
From: ofirgeri at stanford.edu (Ofir Geri)
Date: Tue, 23 Apr 2019 05:48:42 +0000
Subject: [theory-seminar] Theory Seminar (4/26): David Rolnick
Message-ID:
Hi all,
This week's theory seminar talk will be given by David Rolnick from UPenn (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
Complexity of linear regions in deep networks
Speaker: David Rolnick (UPenn)
It is well-known that the expressivity of a neural network depends on its architecture, with deeper networks expressing more complex functions. For ReLU networks, which are piecewise linear, the number of distinct linear regions is a natural measure of expressivity. It is possible to construct networks for which the number of linear regions grows exponentially with depth. However, we show that the expressivity of networks is in practice far below the theoretical maximum. At initialization, we prove that the average number of regions along any one-dimensional subspace grows only linearly, instead of exponentially, in the total number of neurons. More generally, the average number of regions in a k-dimensional subspace is upper bounded by the kth power of the number of neurons, irrespective of network architecture. Our theory and empirical results suggest that this behavior persists during training. We conclude that inductive bias may play a more significant role than expressivity in the success of deep networks. Joint work with Boris Hanin.
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
From wyma at stanford.edu Tue Apr 23 12:01:58 2019
From: wyma at stanford.edu (Weiyun Ma)
Date: Tue, 23 Apr 2019 19:01:58 +0000
Subject: [theory-seminar] Theory Lunch 4/25 -- Shivam Garg
Message-ID:
Hi everyone,
This Thursday at theory lunch, Shivam will tell us about "Sample Amplification: Increasing Dataset Size even when Learning is Impossible." (See abstract below.)
As usual, please join us from noon to 1pm at 463A.
----------------------------------------------------------
Sample Amplification: Increasing Dataset Size even when Learning is Impossible
Speaker: Shivam Garg
Given data drawn from an unknown distribution, D, to what extent is it possible to ``amplify'' this dataset and faithfully output an even larger set of samples that appear to have been drawn from D? We formalize this question as follows: an (n,m) amplification procedure takes as input n independent draws from an unknown distribution D, and outputs a set of m > n ``samples''. An amplification procedure is valid if no algorithm can distinguish the set of m samples produced by the amplifier from a set of m independent draws from D, with probability greater than 2/3. Perhaps surprisingly, in many settings, a valid amplification procedure exists, even in the regime where the size of the input dataset, n, is significantly less than what would be necessary to learn distribution D to non-trivial accuracy. Specifically we consider two fundamental settings: the case where D is an arbitrary discrete distribution supported on k elements, and the case where D is a d-dimensional Gaussian with unknown mean, and fixed covariance matrix. In the first case, we show that an (n, n + O(n/sqrt(k))) amplifier exists. In particular, given n=O(sqrt(k)) samples from D, one can output a set of m=n+1 datapoints, whose total variation distance from the distribution of m i.i.d. draws from D is a small constant, despite the fact that one would need quadratically more data, n=O(k), to learn D up to small constant total variation distance. In the Gaussian case, we show that an (n, n+O(n/sqrt(d))) amplifier exists, even though learning the distribution to small constant total variation distance requires O(d) samples. In both the discrete and Gaussian settings, we show that these results are tight, to constant factors.
Joint work with Brian Axelrod, Vatsal Sharan and Gregory Valiant.
----------------------------------------------------------
Best,
Anna
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
From wyma at stanford.edu Tue Apr 23 12:01:58 2019
From: wyma at stanford.edu (Weiyun Ma)
Date: Tue, 23 Apr 2019 19:01:58 +0000
Subject: [theory-seminar] Theory Lunch 4/25 -- Shivam Garg
Message-ID:
Hi everyone,
This Thursday at theory lunch, Shivam will tell us about "Sample Amplification: Increasing Dataset Size even when Learning is Impossible." (See abstract below.)
As usual, please join us from noon to 1pm at 463A.
----------------------------------------------------------
Sample Amplification: Increasing Dataset Size even when Learning is Impossible
Speaker: Shivam Garg
Given data drawn from an unknown distribution, D, to what extent is it possible to ``amplify'' this dataset and faithfully output an even larger set of samples that appear to have been drawn from D? We formalize this question as follows: an (n,m) amplification procedure takes as input n independent draws from an unknown distribution D, and outputs a set of m > n ``samples''. An amplification procedure is valid if no algorithm can distinguish the set of m samples produced by the amplifier from a set of m independent draws from D, with probability greater than 2/3. Perhaps surprisingly, in many settings, a valid amplification procedure exists, even in the regime where the size of the input dataset, n, is significantly less than what would be necessary to learn distribution D to non-trivial accuracy. Specifically we consider two fundamental settings: the case where D is an arbitrary discrete distribution supported on k elements, and the case where D is a d-dimensional Gaussian with unknown mean, and fixed covariance matrix. In the first case, we show that an (n, n + O(n/sqrt(k))) amplifier exists. In particular, given n=O(sqrt(k)) samples from D, one can output a set of m=n+1 datapoints, whose total variation distance from the distribution of m i.i.d. draws from D is a small constant, despite the fact that one would need quadratically more data, n=O(k), to learn D up to small constant total variation distance. In the Gaussian case, we show that an (n, n+O(n/sqrt(d))) amplifier exists, even though learning the distribution to small constant total variation distance requires O(d) samples. In both the discrete and Gaussian settings, we show that these results are tight, to constant factors.
Joint work with Brian Axelrod, Vatsal Sharan and Gregory Valiant.
----------------------------------------------------------
Best,
Anna
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
From noaj at stanford.edu Thu Apr 25 11:25:02 2019
From: noaj at stanford.edu (Noah Shutty)
Date: Thu, 25 Apr 2019 11:25:02 -0700
Subject: [theory-seminar] Fwd: [sitp_seminar] FYI: SITP Friday group meeting
In-Reply-To:
References: <95BF5476-3794-41A7-9868-7ABD0CD9F140@psiquantum.com>
<25705C72-EB8B-497A-9570-D5FBC214B86C@stanford.edu>
Message-ID:
Naomi Nickerson's talk tomorrow at 10:30 am could be interesting to many of
you.
Naomi will talk about quantum error correction -- description is below and
here is her recent paper:
https://arxiv.org/abs/1810.09621#
Naomi works at PsiQuantum, which is trying to build an optical quantum
computer.
Cheers,
Noah
---------- Forwarded message ---------
From: Julie Shih
Date: Tue, Apr 23, 2019 at 9:50 AM
Subject: [sitp_seminar] FYI: SITP Friday group meeting
To: sitp_seminar at lists.stanford.edu
------------------------------
*From:* Patrick Hayden
Speaker: Naomi Nickerson (PsiQ)
Title: Measurement-based fault tolerance beyond foliation
Friday, 26 April at 10:30am in Varian 355
Abstract:
Quantum error correction is most commonly considered from a ?circuit-based?
point of view, as codes being operated on with measurements. Alternatively
all of quantum error correction can be phrased in the language of
measurement based quantum computing (MBQC) as the construction of fault
tolerant cluster states (FTCSs). While MBQC is often thought of as a
hardware driven choice, it is in fact a useful theoretical tool regardless
of the eventual physical implementation. Surprisingly, while any 2-d
stabilizer code can be used to construct an equivalent FTCS through
foliation, there are FTCSs that cannot be constructed through foliation of
a stabilizer code. I will talk about some examples of this type of cluster
state with remarkable properties.
_______________________________________________
sitp_seminar mailing list
sitp_seminar at lists.stanford.edu
https://mailman.stanford.edu/mailman/listinfo/sitp_seminar
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
From ofirgeri at stanford.edu Fri Apr 26 10:53:39 2019
From: ofirgeri at stanford.edu (Ofir Geri)
Date: Fri, 26 Apr 2019 17:53:39 +0000
Subject: [theory-seminar] Theory Seminar (4/26): David Rolnick
In-Reply-To:
References:
Message-ID:
Reminder: David's talk is today at 3pm.
________________________________
From: Ofir Geri
Sent: Monday, April 22, 2019 10:48 PM
To: thseminar at cs.stanford.edu
Subject: Theory Seminar (4/26): David Rolnick
Hi all,
This week's theory seminar talk will be given by David Rolnick from UPenn (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
Complexity of linear regions in deep networks
Speaker: David Rolnick (UPenn)
It is well-known that the expressivity of a neural network depends on its architecture, with deeper networks expressing more complex functions. For ReLU networks, which are piecewise linear, the number of distinct linear regions is a natural measure of expressivity. It is possible to construct networks for which the number of linear regions grows exponentially with depth. However, we show that the expressivity of networks is in practice far below the theoretical maximum. At initialization, we prove that the average number of regions along any one-dimensional subspace grows only linearly, instead of exponentially, in the total number of neurons. More generally, the average number of regions in a k-dimensional subspace is upper bounded by the kth power of the number of neurons, irrespective of network architecture. Our theory and empirical results suggest that this behavior persists during training. We conclude that inductive bias may play a more significant role than expressivity in the success of deep networks. Joint work with Boris Hanin.
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
From wyma at stanford.edu Mon Apr 29 14:42:56 2019
From: wyma at stanford.edu (Weiyun Ma)
Date: Mon, 29 Apr 2019 21:42:56 +0000
Subject: [theory-seminar] Theory Lunch 5/2 -- Michael Kim
Message-ID:
Hi everyone,
This Thursday at theory lunch, Michael will tell us about "Tracking and Improving Information in the Service of Fairness." (See abstract below.)
As usual, please join us from noon to 1pm at 463A.
----------------------------------------------------------
Tracking and Improving Information in the Service of Fairness
Speaker: Michael Kim
In this talk, we investigate the role of information in *fair* prediction. A common strategy for decision-making uses a predictor to assign individuals a risk score; then, individuals are selected or rejected on the basis of this score. We formalize a framework for measuring the information content of such predictors, showing that increasing information content through a certain kind of "refinement" improves the downstream selection rules across a wide range of fairness measures (e.g. true positive rates, false positive rates, selection rates). In turn, refinements provide a simple but effective tool for reducing disparity in treatment and impact without sacrificing the utility of the predictions. Our results suggest that in many applications, the perceived ?cost of fairness? results from an information disparity across populations, and thus, may be avoided with improved information.
Based on joint work with Sumegha Garg and Omer Reingold.
----------------------------------------------------------
Best,
Anna
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
From wyma at stanford.edu Mon Apr 29 14:42:56 2019
From: wyma at stanford.edu (Weiyun Ma)
Date: Mon, 29 Apr 2019 21:42:56 +0000
Subject: [theory-seminar] Theory Lunch 5/2 -- Michael Kim
Message-ID:
Hi everyone,
This Thursday at theory lunch, Michael will tell us about "Tracking and Improving Information in the Service of Fairness." (See abstract below.)
As usual, please join us from noon to 1pm at 463A.
----------------------------------------------------------
Tracking and Improving Information in the Service of Fairness
Speaker: Michael Kim
In this talk, we investigate the role of information in *fair* prediction. A common strategy for decision-making uses a predictor to assign individuals a risk score; then, individuals are selected or rejected on the basis of this score. We formalize a framework for measuring the information content of such predictors, showing that increasing information content through a certain kind of "refinement" improves the downstream selection rules across a wide range of fairness measures (e.g. true positive rates, false positive rates, selection rates). In turn, refinements provide a simple but effective tool for reducing disparity in treatment and impact without sacrificing the utility of the predictions. Our results suggest that in many applications, the perceived ?cost of fairness? results from an information disparity across populations, and thus, may be avoided with improved information.
Based on joint work with Sumegha Garg and Omer Reingold.
----------------------------------------------------------
Best,
Anna
-------------- next part --------------
An HTML attachment was scrubbed...
URL: