From wajc at stanford.edu Mon Aug 2 08:00:14 2021 From: wajc at stanford.edu (David Wajc) Date: Mon, 2 Aug 2021 18:00:14 +0300 Subject: [theory-seminar] Theory Lunch 08/05: William Kuszmaul (MIT) Message-ID: Hi all, This week's theory lunch will take place Thursday at noon (PDT), at this zoom room . As usual, we'll start with some socializing, followed by a talk starting at 12:30pm. Pras will be MCing. Thanks Pras! Bill will tell us about: *Linear Probing Revisited: Tombstones Mark the Death of Primary Clustering* *Abstract:* First introduced in 1954, linear probing is one of the oldest data structures in computer science, and due to its unrivaled data locality, it continues to be one of the fastest hash tables in practice. It is widely believed and taught, however, that linear probing should never be used at high load factors; this is because primary-clustering effects cause insertions at load factor $1 - 1/x$ to take expected time $\Theta(x^2)$ (rather than the ideal $\Theta(x)$). The dangers of primary clustering, first discovered by Knuth in 1963, have been taught to generations of computer scientists, and have influenced the design of some of many widely used hash tables. We show that primary clustering is not the foregone conclusion that it is reputed to be. We demonstrate that small design decisions in how deletions are implemented have dramatic effects on the asymptotic performance of insertions, so that, even if a hash table operates continuously at a load factor $1 - \Theta(1/x)$, the expected amortized cost per operation is $\tilde{O}(x)$. This is because tombstones created by deletions actually cause an anti-clustering effect that combats primary clustering. We also present a new variant of linear probing (which we call graveyard hashing) that completely eliminates primary clustering on any sequence of operations: if, when an operation is performed, the current load factor is $1 - 1/x$ for some $x$, then the expected cost of the operation is $O(x)$. One corollary is that, in the external-memory model with a data blocks of size $B$, graveyard hashing offers the following remarkable guarantee: at any load factor $1-1/x$ satisfying $x=o(B)$, graveyard hashing achieves $1+o(1)$ expected block transfers per operation. Past external-memory hash tables have only been able to offer a $1+o(1)$ guarantee when the block size $B$ is at least $\Omega(x^2)$. Based on joint work with Michael A. Bender, Bradley C. Kuszmaul Cheers, David -------------- next part -------------- An HTML attachment was scrubbed... URL: From wajc at stanford.edu Mon Aug 2 08:00:14 2021 From: wajc at stanford.edu (David Wajc) Date: Mon, 2 Aug 2021 18:00:14 +0300 Subject: [theory-seminar] Theory Lunch 08/05: William Kuszmaul (MIT) Message-ID: Hi all, This week's theory lunch will take place Thursday at noon (PDT), at this zoom room . As usual, we'll start with some socializing, followed by a talk starting at 12:30pm. Pras will be MCing. Thanks Pras! Bill will tell us about: *Linear Probing Revisited: Tombstones Mark the Death of Primary Clustering* *Abstract:* First introduced in 1954, linear probing is one of the oldest data structures in computer science, and due to its unrivaled data locality, it continues to be one of the fastest hash tables in practice. It is widely believed and taught, however, that linear probing should never be used at high load factors; this is because primary-clustering effects cause insertions at load factor $1 - 1/x$ to take expected time $\Theta(x^2)$ (rather than the ideal $\Theta(x)$). The dangers of primary clustering, first discovered by Knuth in 1963, have been taught to generations of computer scientists, and have influenced the design of some of many widely used hash tables. We show that primary clustering is not the foregone conclusion that it is reputed to be. We demonstrate that small design decisions in how deletions are implemented have dramatic effects on the asymptotic performance of insertions, so that, even if a hash table operates continuously at a load factor $1 - \Theta(1/x)$, the expected amortized cost per operation is $\tilde{O}(x)$. This is because tombstones created by deletions actually cause an anti-clustering effect that combats primary clustering. We also present a new variant of linear probing (which we call graveyard hashing) that completely eliminates primary clustering on any sequence of operations: if, when an operation is performed, the current load factor is $1 - 1/x$ for some $x$, then the expected cost of the operation is $O(x)$. One corollary is that, in the external-memory model with a data blocks of size $B$, graveyard hashing offers the following remarkable guarantee: at any load factor $1-1/x$ satisfying $x=o(B)$, graveyard hashing achieves $1+o(1)$ expected block transfers per operation. Past external-memory hash tables have only been able to offer a $1+o(1)$ guarantee when the block size $B$ is at least $\Omega(x^2)$. Based on joint work with Michael A. Bender, Bradley C. Kuszmaul Cheers, David -------------- next part -------------- An HTML attachment was scrubbed... URL: From wajc at stanford.edu Tue Aug 3 08:28:26 2021 From: wajc at stanford.edu (David Wajc) Date: Tue, 3 Aug 2021 18:28:26 +0300 Subject: [theory-seminar] More mentorship resources Message-ID: Members of the group may find this learning theory alliance workshop (Aug 4 + 5)?focussed on grad school admissions, fellowships and academic jobs?relevant. (This is not exclusive to ML-theory; speakers are from broad areas of TCS, and the advice should be broadly relevant.) Best, David -------------- next part -------------- An HTML attachment was scrubbed... URL: From wajc at stanford.edu Tue Aug 3 08:28:26 2021 From: wajc at stanford.edu (David Wajc) Date: Tue, 3 Aug 2021 18:28:26 +0300 Subject: [theory-seminar] More mentorship resources Message-ID: Members of the group may find this learning theory alliance workshop (Aug 4 + 5)?focussed on grad school admissions, fellowships and academic jobs?relevant. (This is not exclusive to ML-theory; speakers are from broad areas of TCS, and the advice should be broadly relevant.) Best, David -------------- next part -------------- An HTML attachment was scrubbed... URL: From pras1712 at stanford.edu Thu Aug 5 11:41:51 2021 From: pras1712 at stanford.edu (Prasanna Ramakrishnan) Date: Thu, 5 Aug 2021 14:41:51 -0400 Subject: [theory-seminar] Theory Lunch 08/05: William Kuszmaul (MIT) In-Reply-To: References: Message-ID: Hey everyone! Gentle reminder that theory lunch is starting in about 18 minutes, with the talk starting half an hour later. Hope to see you all there! On Mon, Aug 2, 2021 at 11:00 AM David Wajc wrote: > Hi all, > > This week's theory lunch will take place Thursday at noon (PDT), at this > zoom room > . As > usual, we'll start with some socializing, followed by a talk starting at > 12:30pm. Pras will be MCing. Thanks Pras! > Bill will tell us about: *Linear Probing Revisited: Tombstones Mark the > Death of Primary Clustering* > > *Abstract:* First introduced in 1954, linear probing is one of the oldest > data structures in computer science, and due to its unrivaled data > locality, it continues to be one of the fastest hash tables in practice. It > is widely believed and taught, however, that linear probing should never be > used at high load factors; this is because primary-clustering effects cause > insertions at load factor $1 - 1/x$ to take expected time $\Theta(x^2)$ > (rather than the ideal $\Theta(x)$). The dangers of primary clustering, > first discovered by Knuth in 1963, have been taught to generations of > computer scientists, and have influenced the design of some of many widely > used hash tables. > > We show that primary clustering is not the foregone conclusion that it is > reputed to be. We demonstrate that small design decisions in how deletions > are implemented have dramatic effects on the asymptotic performance of > insertions, so that, even if a hash table operates continuously at a load > factor $1 - \Theta(1/x)$, the expected amortized cost per operation is > $\tilde{O}(x)$. This is because tombstones created by deletions actually > cause an anti-clustering effect that combats primary clustering. > > We also present a new variant of linear probing (which we call graveyard > hashing) that completely eliminates primary clustering on any sequence of > operations: if, when an operation is performed, the current load factor is > $1 - 1/x$ for some $x$, then the expected cost of the operation is $O(x)$. > One corollary is that, in the external-memory model with a data blocks of > size $B$, graveyard hashing offers the following remarkable guarantee: at > any load factor $1-1/x$ satisfying $x=o(B)$, graveyard hashing achieves > $1+o(1)$ expected block transfers per operation. Past external-memory hash > tables have only been able to offer a $1+o(1)$ guarantee when the block > size $B$ is at least $\Omega(x^2)$. > > Based on joint work with Michael A. Bender, Bradley C. Kuszmaul > > Cheers, David > -------------- next part -------------- An HTML attachment was scrubbed... URL: From pras1712 at stanford.edu Thu Aug 5 11:41:51 2021 From: pras1712 at stanford.edu (Prasanna Ramakrishnan) Date: Thu, 5 Aug 2021 14:41:51 -0400 Subject: [theory-seminar] Theory Lunch 08/05: William Kuszmaul (MIT) In-Reply-To: References: Message-ID: Hey everyone! Gentle reminder that theory lunch is starting in about 18 minutes, with the talk starting half an hour later. Hope to see you all there! On Mon, Aug 2, 2021 at 11:00 AM David Wajc wrote: > Hi all, > > This week's theory lunch will take place Thursday at noon (PDT), at this > zoom room > . As > usual, we'll start with some socializing, followed by a talk starting at > 12:30pm. Pras will be MCing. Thanks Pras! > Bill will tell us about: *Linear Probing Revisited: Tombstones Mark the > Death of Primary Clustering* > > *Abstract:* First introduced in 1954, linear probing is one of the oldest > data structures in computer science, and due to its unrivaled data > locality, it continues to be one of the fastest hash tables in practice. It > is widely believed and taught, however, that linear probing should never be > used at high load factors; this is because primary-clustering effects cause > insertions at load factor $1 - 1/x$ to take expected time $\Theta(x^2)$ > (rather than the ideal $\Theta(x)$). The dangers of primary clustering, > first discovered by Knuth in 1963, have been taught to generations of > computer scientists, and have influenced the design of some of many widely > used hash tables. > > We show that primary clustering is not the foregone conclusion that it is > reputed to be. We demonstrate that small design decisions in how deletions > are implemented have dramatic effects on the asymptotic performance of > insertions, so that, even if a hash table operates continuously at a load > factor $1 - \Theta(1/x)$, the expected amortized cost per operation is > $\tilde{O}(x)$. This is because tombstones created by deletions actually > cause an anti-clustering effect that combats primary clustering. > > We also present a new variant of linear probing (which we call graveyard > hashing) that completely eliminates primary clustering on any sequence of > operations: if, when an operation is performed, the current load factor is > $1 - 1/x$ for some $x$, then the expected cost of the operation is $O(x)$. > One corollary is that, in the external-memory model with a data blocks of > size $B$, graveyard hashing offers the following remarkable guarantee: at > any load factor $1-1/x$ satisfying $x=o(B)$, graveyard hashing achieves > $1+o(1)$ expected block transfers per operation. Past external-memory hash > tables have only been able to offer a $1+o(1)$ guarantee when the block > size $B$ is at least $\Omega(x^2)$. > > Based on joint work with Michael A. Bender, Bradley C. Kuszmaul > > Cheers, David > -------------- next part -------------- An HTML attachment was scrubbed... URL: From wajc at stanford.edu Mon Aug 9 10:08:41 2021 From: wajc at stanford.edu (David Wajc) Date: Mon, 9 Aug 2021 20:08:41 +0300 Subject: [theory-seminar] Theory Lunch 08/12: Mingda Qiao Message-ID: Hi all, This week's theory lunch will take place Thursday at noon (PDT), at this zoom room . As usual, we'll start with some socializing, followed by a talk starting at 12:30pm. Pras will be MCing. Thanks Pras! Mingda will tell us about: *Exponential Weights Algorithms for Selective Learning* *Abstract:* We study the selective learning problem, in which a learner observes $n$ labeled data points one at a time. At a time of its choosing, the learner picks a window length $w$ and a hypothesis $h$ from a given hypothesis class $H$, and then labels the next $w$ data points using $h$. The excess risk incurred by the learner is defined as the difference between the average loss of $h$ over the $w$ data points and the smallest possible average loss among all hypotheses in $H$ over those data points. We give an improved algorithm, termed the hybrid exponential weights algorithm, that achieves an expected excess risk of $O((\log\log|H| + \log\log n)/\log n)$. This gives a doubly exponential improvement in the dependence on $|H|$ over the best known upper bound. We also prove an almost matching lower bound, showing that the $\log\log|H|$ dependence is necessary. Based on joint work with Greg Valiant. Cheers, David -------------- next part -------------- An HTML attachment was scrubbed... URL: From wajc at stanford.edu Mon Aug 9 10:08:41 2021 From: wajc at stanford.edu (David Wajc) Date: Mon, 9 Aug 2021 20:08:41 +0300 Subject: [theory-seminar] Theory Lunch 08/12: Mingda Qiao Message-ID: Hi all, This week's theory lunch will take place Thursday at noon (PDT), at this zoom room . As usual, we'll start with some socializing, followed by a talk starting at 12:30pm. Pras will be MCing. Thanks Pras! Mingda will tell us about: *Exponential Weights Algorithms for Selective Learning* *Abstract:* We study the selective learning problem, in which a learner observes $n$ labeled data points one at a time. At a time of its choosing, the learner picks a window length $w$ and a hypothesis $h$ from a given hypothesis class $H$, and then labels the next $w$ data points using $h$. The excess risk incurred by the learner is defined as the difference between the average loss of $h$ over the $w$ data points and the smallest possible average loss among all hypotheses in $H$ over those data points. We give an improved algorithm, termed the hybrid exponential weights algorithm, that achieves an expected excess risk of $O((\log\log|H| + \log\log n)/\log n)$. This gives a doubly exponential improvement in the dependence on $|H|$ over the best known upper bound. We also prove an almost matching lower bound, showing that the $\log\log|H|$ dependence is necessary. Based on joint work with Greg Valiant. Cheers, David -------------- next part -------------- An HTML attachment was scrubbed... URL: From wajc at stanford.edu Wed Aug 11 10:06:03 2021 From: wajc at stanford.edu (David Wajc) Date: Wed, 11 Aug 2021 20:06:03 +0300 Subject: [theory-seminar] Fwd: WALDO 2021 In-Reply-To: References: Message-ID: FYI, a workshop of possible interest to members of the theory group, covering such topics as streaming, property testing, and sublinear algorithms, August 23-25. Registration is free, but required by August 20. Cheers, David ---------- Forwarded message --------- From: Ainesh Bakshi Date: Tue, 10 Aug 2021 at 22:55 Subject: WALDO 2021 To: David Wajc Hi everyone, We would like to invite you to the Workshop on Algorithms for Large Data (Online), which we're organizing from Monday, August 23rd to Wednesday, August 25th. Our goal for the workshop is to generate new collaborations through an emphasis on open problems in sublinear algorithms and numerical linear algebra. We're welcoming a fantastic set of speakers and there will also be networking events such as junior-senior mentoring and a poster session. Everything is online, registration is free (though necessary to estimate constraints for the virtual space), and everyone is welcome. More details are available at https://waldo2021.github.io/ We hope to see you there! Thanks and best regards, Co-organizers, Ainesh Bakshi Rajesh Jayaram Samson Zhou -------------- next part -------------- An HTML attachment was scrubbed... URL: From wajc at stanford.edu Wed Aug 11 10:06:03 2021 From: wajc at stanford.edu (David Wajc) Date: Wed, 11 Aug 2021 20:06:03 +0300 Subject: [theory-seminar] Fwd: WALDO 2021 In-Reply-To: References: Message-ID: FYI, a workshop of possible interest to members of the theory group, covering such topics as streaming, property testing, and sublinear algorithms, August 23-25. Registration is free, but required by August 20. Cheers, David ---------- Forwarded message --------- From: Ainesh Bakshi Date: Tue, 10 Aug 2021 at 22:55 Subject: WALDO 2021 To: David Wajc Hi everyone, We would like to invite you to the Workshop on Algorithms for Large Data (Online), which we're organizing from Monday, August 23rd to Wednesday, August 25th. Our goal for the workshop is to generate new collaborations through an emphasis on open problems in sublinear algorithms and numerical linear algebra. We're welcoming a fantastic set of speakers and there will also be networking events such as junior-senior mentoring and a poster session. Everything is online, registration is free (though necessary to estimate constraints for the virtual space), and everyone is welcome. More details are available at https://waldo2021.github.io/ We hope to see you there! Thanks and best regards, Co-organizers, Ainesh Bakshi Rajesh Jayaram Samson Zhou -------------- next part -------------- An HTML attachment was scrubbed... URL: From pras1712 at stanford.edu Thu Aug 12 11:56:45 2021 From: pras1712 at stanford.edu (Prasanna Ramakrishnan) Date: Thu, 12 Aug 2021 14:56:45 -0400 Subject: [theory-seminar] Theory Lunch 08/12: Mingda Qiao In-Reply-To: References: Message-ID: Reminder: Theory lunch starts in a few minutes, with the talk starting half an hour later! See you all there! On Mon, Aug 9, 2021 at 1:09 PM David Wajc wrote: > Hi all, > > This week's theory lunch will take place Thursday at noon (PDT), at this > zoom room > . As > usual, we'll start with some socializing, followed by a talk starting at > 12:30pm. Pras will be MCing. Thanks Pras! > Mingda will tell us about: *Exponential Weights Algorithms for Selective > Learning* > > *Abstract:* We study the selective learning problem, in which a learner > observes $n$ labeled data points one at a time. At a time of its choosing, > the learner picks a window length $w$ and a hypothesis $h$ from a given > hypothesis class $H$, and then labels the next $w$ data points using $h$. > The excess risk incurred by the learner is defined as the difference > between the average loss of $h$ over the $w$ data points and the smallest > possible average loss among all hypotheses in $H$ over those data points. > > We give an improved algorithm, termed the hybrid exponential weights > algorithm, that achieves an expected excess risk of $O((\log\log|H| + > \log\log n)/\log n)$. This gives a doubly exponential improvement in the > dependence on $|H|$ over the best known upper bound. We also prove an > almost matching lower bound, showing that the $\log\log|H|$ dependence is > necessary. > > Based on joint work with Greg Valiant. > > Cheers, David > -------------- next part -------------- An HTML attachment was scrubbed... URL: From pras1712 at stanford.edu Thu Aug 12 11:56:45 2021 From: pras1712 at stanford.edu (Prasanna Ramakrishnan) Date: Thu, 12 Aug 2021 14:56:45 -0400 Subject: [theory-seminar] Theory Lunch 08/12: Mingda Qiao In-Reply-To: References: Message-ID: Reminder: Theory lunch starts in a few minutes, with the talk starting half an hour later! See you all there! On Mon, Aug 9, 2021 at 1:09 PM David Wajc wrote: > Hi all, > > This week's theory lunch will take place Thursday at noon (PDT), at this > zoom room > . As > usual, we'll start with some socializing, followed by a talk starting at > 12:30pm. Pras will be MCing. Thanks Pras! > Mingda will tell us about: *Exponential Weights Algorithms for Selective > Learning* > > *Abstract:* We study the selective learning problem, in which a learner > observes $n$ labeled data points one at a time. At a time of its choosing, > the learner picks a window length $w$ and a hypothesis $h$ from a given > hypothesis class $H$, and then labels the next $w$ data points using $h$. > The excess risk incurred by the learner is defined as the difference > between the average loss of $h$ over the $w$ data points and the smallest > possible average loss among all hypotheses in $H$ over those data points. > > We give an improved algorithm, termed the hybrid exponential weights > algorithm, that achieves an expected excess risk of $O((\log\log|H| + > \log\log n)/\log n)$. This gives a doubly exponential improvement in the > dependence on $|H|$ over the best known upper bound. We also prove an > almost matching lower bound, showing that the $\log\log|H|$ dependence is > necessary. > > Based on joint work with Greg Valiant. > > Cheers, David > -------------- next part -------------- An HTML attachment was scrubbed... URL: From marykw at stanford.edu Fri Aug 13 15:37:40 2021 From: marykw at stanford.edu (Mary Wootters) Date: Fri, 13 Aug 2021 16:37:40 -0600 Subject: [theory-seminar] Call for participation: APPROX/RANDOM 2021 Message-ID: Hi all, The 25th International Workshop on Randomization and Computation (RANDOM 2021) and the 24th International Workshop on Approximation Algorithms for Combinatorial Optimization Problems (APPROX 2021) will be starting on Monday August 16! The conferences will be held as parallel virtual conferences, August 16-18, 2021. RANDOM 2021 focuses on applications of randomness to computational and combinatorial problems while APPROX 2021 focuses on algorithmic and complexity theoretic issues relevant to the development of efficient approximate solutions to computationally difficult problems. To learn more about the conferences and program, visit: APPROX: https://approxconference.wordpress.com/approx-2021/ RANDOM: https://randomconference.com/random-2021-home/ In addition to an exciting program of live talks and discussion (to complement pre-recorded talks), the conference will feature two invited talks, by Jelani Nelson (UC Berkeley) and Vera Traub (ETH Zurich), as well as a social event with trivia and a cartoon caption contest! Registration is only $10 for general audience members. You can register here: https://www.eventbrite.com/e/approx-2021-and-random-2021-tickets-162840998811?discount=audience Hope to see you at the conference! Mary -- Mary Wootters (she/her) Assistant Professor of Computer Science and Electrical Engineering Stanford University -------------- next part -------------- An HTML attachment was scrubbed... URL: From wajc at stanford.edu Mon Aug 16 11:40:20 2021 From: wajc at stanford.edu (David Wajc) Date: Mon, 16 Aug 2021 11:40:20 -0700 Subject: [theory-seminar] Theory Lunch 08/19: Shashwat Silas (Google) Message-ID: Hi all, This week's theory lunch will take place Thursday at noon (PDT), at this zoom room . As usual, we'll start with some socializing, followed by a talk starting at 12:30pm. Shashwat will tell us about: *Real-time decoding of rateless codes* *Abstract:* Rateless codes have been widely studied in academia and are used in several production software systems for networking. There are still many interesting open problems to consider. We will have a brief introduction to rateless codes and look at the problem of real-time oblivious decoding of rateless codes. Rateless codes are used for transmission of information across channels whose rate of erasure is unknown. In such a code, an infinite stream of encoding symbols can be generated from the message and sent across the erasure channel, and the decoder can decode the message after it has successfully collected a certain number of encoding symbols. A rateless erasure code is real-time oblivious if rather than collecting encoding symbols as they are received, the receiver either immediately decodes or discards each symbol it receives. Efficient real-time oblivious erasure correction uses a feedback channel in order to maximize the probability that a received encoding symbol is decoded rather than discarded. We construct codes which are real-time oblivious, but require fewer feedback messages and have faster decoding compared to previous work. Specifically, for a message of length *k*, we improve the expected complexity of the feedback channel from O(k^(1/2)) to O(1), and the expected decoding complexity from O(klog(k)) to O(k) Our method involves using an appropriate block erasure code to first encode the *k *message symbols, and then using a truncated version of the real-time oblivious erasure correction of Beimel et al (2007) to transmit the encoded message to the receiver, which then uses the decoding algorithm for the outer code to recover the message. Finally, we will look at simulations of non-oblivious models of real-time decoding which have even better real-time decoding properties. Cheers, David -------------- next part -------------- An HTML attachment was scrubbed... URL: From wajc at stanford.edu Mon Aug 16 11:40:20 2021 From: wajc at stanford.edu (David Wajc) Date: Mon, 16 Aug 2021 11:40:20 -0700 Subject: [theory-seminar] Theory Lunch 08/19: Shashwat Silas (Google) Message-ID: Hi all, This week's theory lunch will take place Thursday at noon (PDT), at this zoom room . As usual, we'll start with some socializing, followed by a talk starting at 12:30pm. Shashwat will tell us about: *Real-time decoding of rateless codes* *Abstract:* Rateless codes have been widely studied in academia and are used in several production software systems for networking. There are still many interesting open problems to consider. We will have a brief introduction to rateless codes and look at the problem of real-time oblivious decoding of rateless codes. Rateless codes are used for transmission of information across channels whose rate of erasure is unknown. In such a code, an infinite stream of encoding symbols can be generated from the message and sent across the erasure channel, and the decoder can decode the message after it has successfully collected a certain number of encoding symbols. A rateless erasure code is real-time oblivious if rather than collecting encoding symbols as they are received, the receiver either immediately decodes or discards each symbol it receives. Efficient real-time oblivious erasure correction uses a feedback channel in order to maximize the probability that a received encoding symbol is decoded rather than discarded. We construct codes which are real-time oblivious, but require fewer feedback messages and have faster decoding compared to previous work. Specifically, for a message of length *k*, we improve the expected complexity of the feedback channel from O(k^(1/2)) to O(1), and the expected decoding complexity from O(klog(k)) to O(k) Our method involves using an appropriate block erasure code to first encode the *k *message symbols, and then using a truncated version of the real-time oblivious erasure correction of Beimel et al (2007) to transmit the encoded message to the receiver, which then uses the decoding algorithm for the outer code to recover the message. Finally, we will look at simulations of non-oblivious models of real-time decoding which have even better real-time decoding properties. Cheers, David -------------- next part -------------- An HTML attachment was scrubbed... URL: From sidford at stanford.edu Tue Aug 17 21:13:38 2021 From: sidford at stanford.edu (Aaron Sidford) Date: Tue, 17 Aug 2021 22:13:38 -0600 Subject: [theory-seminar] Fwd: Registration now open: 1st ACM Conference on Equity and Access in Algorithms, Mechanisms, and Optimization (EAAMO'21) In-Reply-To: References: Message-ID: Hi all, This might be of interest to many of you. All the best, Aaron ---------- Forwarded message --------- From: Irene Yuan Lo Date: Tue, Aug 17, 2021 at 6:24 PM Subject: [MSandE Faculty] Registration now open: 1st ACM Conference on Equity and Access in Algorithms, Mechanisms, and Optimization (EAAMO'21) To: socialalgorithms at lists.stanford.edu , internetalgs at lists.stanford.edu , econ-marketdesign at lists.stanford.edu , msande-phd at lists.stanford.edu , msande-acfaculty at lists.stanford.edu * Hi everyone, As many of you may know, I've been part of launching a new ACM conference series and publication venue: the 1st ACM Conference on Equity and Access in Algorithms, Mechanisms, and Optimization (EAAMO'21) . Registration is now live, and I'd be delighted to see many of you there! See below for more details. Irene *** We are thrilled to announce that the registration for EAAMO ?21 is now live! Please register for regular admission on Eventbrite by September 10, 2021. Conference registration is$20 for ACM members, $15 for students, and$35 for non-ACM members. We also provide financial assistance and data grants in order to waive registration fees and provide data plans to facilitate virtual attendance. Please apply here before September 10, 2021. A main goal of the conference is to bridge research and practice. Please nominate practitioners working with underserved and disadvantaged communities to join us at the conference (you can also nominate yourself if you are a practitioner). Invited practitioners will be included in facilitated discussions with researchers. For more information, please see below or visit our website and contact us at gc at eaamo.org with any questions. *** The inaugural Conference on Equity and Access in Algorithms, Mechanisms, and Optimization (EAAMO ?21) will take place on October 5-9, 2021, virtually, on Zoom and Gather.town. EAAMO ?21 will be sponsored by ACM SIGAI and SIGecom . The goal of this event is to highlight work where techniques from algorithms, optimization, and mechanism design, along with insights from the social sciences and humanistic studies, can improve access to opportunity for historically underserved and disadvantaged communities. The conference aims to foster a multi-disciplinary community, facilitating interactions between academia, industry, and the public and voluntary sectors. The program will feature keynote presentations from researchers and practitioners as well as contributed presentations in the research and policy & practice tracks. We are excited to host a series of keynote speakers from a variety of fields: Solomon Assefa (IBM Research), Dirk Bergemann (Yale University), Ellora Derenoncourt (University of California, Berkeley), Ashish Goel (Stanford University), Mary Gray (Microsoft Research), Krishna Gummadi (Max Planck Institute for Software Systems), Avinatan Hassidim (Bar Ilan University), Radhika Khosla (University of Oxford), Sylvia Ortega Salazar (National College of Vocational and Professional Training), and Trooper Sanders (Benefits Data Trust). ACM EAAMO is part of the Mechanism Design for Social Good (MD4SG) initiative, and builds on the MD4SG technical workshop series and tutorials at conferences including ACM EC, ACM COMPASS, ACM FAccT, and WINE. * -- Irene Lo Assistant Professor in Management Science & Engineering Stanford University ilo at stanford.edu | 909-859-4183 Website: https://sites.google.com/view/irene-lo -------------- next part -------------- An HTML attachment was scrubbed... URL: From sidford at stanford.edu Tue Aug 17 21:13:38 2021 From: sidford at stanford.edu (Aaron Sidford) Date: Tue, 17 Aug 2021 22:13:38 -0600 Subject: [theory-seminar] Fwd: Registration now open: 1st ACM Conference on Equity and Access in Algorithms, Mechanisms, and Optimization (EAAMO'21) In-Reply-To: References: Message-ID: Hi all, This might be of interest to many of you. All the best, Aaron ---------- Forwarded message --------- From: Irene Yuan Lo Date: Tue, Aug 17, 2021 at 6:24 PM Subject: [MSandE Faculty] Registration now open: 1st ACM Conference on Equity and Access in Algorithms, Mechanisms, and Optimization (EAAMO'21) To: socialalgorithms at lists.stanford.edu , internetalgs at lists.stanford.edu , econ-marketdesign at lists.stanford.edu , msande-phd at lists.stanford.edu , msande-acfaculty at lists.stanford.edu * Hi everyone, As many of you may know, I've been part of launching a new ACM conference series and publication venue: the 1st ACM Conference on Equity and Access in Algorithms, Mechanisms, and Optimization (EAAMO'21) . Registration is now live, and I'd be delighted to see many of you there! See below for more details. Irene *** We are thrilled to announce that the registration for EAAMO ?21 is now live! Please register for regular admission on Eventbrite by September 10, 2021. Conference registration is $20 for ACM members,$15 for students, and \$35 for non-ACM members. We also provide financial assistance and data grants in order to waive registration fees and provide data plans to facilitate virtual attendance. Please apply here before September 10, 2021. A main goal of the conference is to bridge research and practice. Please nominate practitioners working with underserved and disadvantaged communities to join us at the conference (you can also nominate yourself if you are a practitioner). Invited practitioners will be included in facilitated discussions with researchers. For more information, please see below or visit our website and contact us at gc at eaamo.org with any questions. *** The inaugural Conference on Equity and Access in Algorithms, Mechanisms, and Optimization (EAAMO ?21) will take place on October 5-9, 2021, virtually, on Zoom and Gather.town. EAAMO ?21 will be sponsored by ACM SIGAI and SIGecom . The goal of this event is to highlight work where techniques from algorithms, optimization, and mechanism design, along with insights from the social sciences and humanistic studies, can improve access to opportunity for historically underserved and disadvantaged communities. The conference aims to foster a multi-disciplinary community, facilitating interactions between academia, industry, and the public and voluntary sectors. The program will feature keynote presentations from researchers and practitioners as well as contributed presentations in the research and policy & practice tracks. We are excited to host a series of keynote speakers from a variety of fields: Solomon Assefa (IBM Research), Dirk Bergemann (Yale University), Ellora Derenoncourt (University of California, Berkeley), Ashish Goel (Stanford University), Mary Gray (Microsoft Research), Krishna Gummadi (Max Planck Institute for Software Systems), Avinatan Hassidim (Bar Ilan University), Radhika Khosla (University of Oxford), Sylvia Ortega Salazar (National College of Vocational and Professional Training), and Trooper Sanders (Benefits Data Trust). ACM EAAMO is part of the Mechanism Design for Social Good (MD4SG) initiative, and builds on the MD4SG technical workshop series and tutorials at conferences including ACM EC, ACM COMPASS, ACM FAccT, and WINE. * -- Irene Lo Assistant Professor in Management Science & Engineering Stanford University ilo at stanford.edu | 909-859-4183 Website: https://sites.google.com/view/irene-lo -------------- next part -------------- An HTML attachment was scrubbed... URL: From wajc at stanford.edu Thu Aug 19 09:05:51 2021 From: wajc at stanford.edu (David Wajc) Date: Thu, 19 Aug 2021 09:05:51 -0700 Subject: [theory-seminar] Theory Lunch 08/19: Shashwat Silas (Google) In-Reply-To: References: Message-ID: Gentle reminder: theory lunch starting in three hours, with talk starting half an hour after that. See you there! Cheers, David On Mon, 16 Aug 2021 at 11:40, David Wajc wrote: > Hi all, > > This week's theory lunch will take place Thursday at noon (PDT), at this > zoom room > . As > usual, we'll start with some socializing, followed by a talk starting at > 12:30pm. > Shashwat will tell us about: *Real-time decoding of rateless codes* > > *Abstract:* Rateless codes have been widely studied in academia and are > used in several production software systems for networking. There are still > many interesting open problems to consider. We will have a brief > introduction to rateless codes and look at the problem of real-time > oblivious decoding of rateless codes. > > Rateless codes are used for transmission of information across channels > whose rate of erasure is unknown. In such a code, an infinite stream of > encoding symbols can be generated from the message and sent across the > erasure channel, and the decoder can decode the message after it has > successfully collected a certain number of encoding symbols. A rateless > erasure code is real-time oblivious if rather than collecting encoding > symbols as they are received, the receiver either immediately decodes or > discards each symbol it receives. > > Efficient real-time oblivious erasure correction uses a feedback channel > in order to maximize the probability that a received encoding symbol is > decoded rather than discarded. We construct codes which are real-time > oblivious, but require fewer feedback messages and have faster decoding > compared to previous work. Specifically, for a message of length *k*, we > improve the expected complexity of the feedback channel from O(k^(1/2)) to > O(1), and the expected decoding complexity from O(klog(k)) to O(k) Our > method involves using an appropriate block erasure code to first encode the > *k *message symbols, and then using a truncated version of the real-time > oblivious erasure correction of Beimel et al (2007) to transmit the encoded > message to the receiver, which then uses the decoding algorithm for the > outer code to recover the message. > > Finally, we will look at simulations of non-oblivious models of real-time > decoding which have even better real-time decoding properties. > > Cheers, > David > -------------- next part -------------- An HTML attachment was scrubbed... URL: From wajc at stanford.edu Thu Aug 19 09:05:51 2021 From: wajc at stanford.edu (David Wajc) Date: Thu, 19 Aug 2021 09:05:51 -0700 Subject: [theory-seminar] Theory Lunch 08/19: Shashwat Silas (Google) In-Reply-To: References: Message-ID: Gentle reminder: theory lunch starting in three hours, with talk starting half an hour after that. See you there! Cheers, David On Mon, 16 Aug 2021 at 11:40, David Wajc wrote: > Hi all, > > This week's theory lunch will take place Thursday at noon (PDT), at this > zoom room > . As > usual, we'll start with some socializing, followed by a talk starting at > 12:30pm. > Shashwat will tell us about: *Real-time decoding of rateless codes* > > *Abstract:* Rateless codes have been widely studied in academia and are > used in several production software systems for networking. There are still > many interesting open problems to consider. We will have a brief > introduction to rateless codes and look at the problem of real-time > oblivious decoding of rateless codes. > > Rateless codes are used for transmission of information across channels > whose rate of erasure is unknown. In such a code, an infinite stream of > encoding symbols can be generated from the message and sent across the > erasure channel, and the decoder can decode the message after it has > successfully collected a certain number of encoding symbols. A rateless > erasure code is real-time oblivious if rather than collecting encoding > symbols as they are received, the receiver either immediately decodes or > discards each symbol it receives. > > Efficient real-time oblivious erasure correction uses a feedback channel > in order to maximize the probability that a received encoding symbol is > decoded rather than discarded. We construct codes which are real-time > oblivious, but require fewer feedback messages and have faster decoding > compared to previous work. Specifically, for a message of length *k*, we > improve the expected complexity of the feedback channel from O(k^(1/2)) to > O(1), and the expected decoding complexity from O(klog(k)) to O(k) Our > method involves using an appropriate block erasure code to first encode the > *k *message symbols, and then using a truncated version of the real-time > oblivious erasure correction of Beimel et al (2007) to transmit the encoded > message to the receiver, which then uses the decoding algorithm for the > outer code to recover the message. > > Finally, we will look at simulations of non-oblivious models of real-time > decoding which have even better real-time decoding properties. > > Cheers, > David > -------------- next part -------------- An HTML attachment was scrubbed... URL: From wajc at stanford.edu Mon Aug 23 13:12:20 2021 From: wajc at stanford.edu (David Wajc) Date: Mon, 23 Aug 2021 13:12:20 -0700 Subject: [theory-seminar] Theory Lunch 08/26: Yujia Jin Message-ID: Hi all, The last theory lunch of the summer quarter will take place Thursday at noon (PDT), at this zoom room . As usual, we'll start with some socializing, followed by a talk starting at 12:30pm. Yujia will tell us about: *Thinking Inside the Ball: Near-Optimal Minimization of the Maximal Loss* *Abstract:* In this work, we characterize the complexity of minimizing the maximal of N functions for convex, Lipschitz functions f_1,? , f_N. For non-smooth functions, existing methods require O(N*\epsilon^{-2}) queries to a first-order oracle to compute an \epsilon-suboptimal point and \Otil{N*\epsilon^{-1}} queries if the functions f_i are O(1/\epsilon)-smooth. We develop methods with improved complexity bounds of \Otil{N*\epsilon^{-2/3} + \epsilon^{-8/3}} in the non-smooth case and \Otil{N*\epsilon^{-2/3} + \sqrt{N}*\epsilon^{-1}} in the O(1/\epsilon)-smooth case. Our methods consist of a recently proposed ball optimization oracle acceleration algorithm (which we refine) and a careful implementation of said oracle for the softmax function. We also prove an oracle complexity lower bound scaling as \Omega(N*\epsilon^{-2/3}), showing that our dependence on N is optimal up to polylogarithmic factors. This talk is based on joint work with Yair Carmon, Arun Jambulapati, and Aaron Sidford. Cheers, David -------------- next part -------------- An HTML attachment was scrubbed... URL: From wajc at stanford.edu Mon Aug 23 13:12:20 2021 From: wajc at stanford.edu (David Wajc) Date: Mon, 23 Aug 2021 13:12:20 -0700 Subject: [theory-seminar] Theory Lunch 08/26: Yujia Jin Message-ID: Hi all, The last theory lunch of the summer quarter will take place Thursday at noon (PDT), at this zoom room . As usual, we'll start with some socializing, followed by a talk starting at 12:30pm. Yujia will tell us about: *Thinking Inside the Ball: Near-Optimal Minimization of the Maximal Loss* *Abstract:* In this work, we characterize the complexity of minimizing the maximal of N functions for convex, Lipschitz functions f_1,? , f_N. For non-smooth functions, existing methods require O(N*\epsilon^{-2}) queries to a first-order oracle to compute an \epsilon-suboptimal point and \Otil{N*\epsilon^{-1}} queries if the functions f_i are O(1/\epsilon)-smooth. We develop methods with improved complexity bounds of \Otil{N*\epsilon^{-2/3} + \epsilon^{-8/3}} in the non-smooth case and \Otil{N*\epsilon^{-2/3} + \sqrt{N}*\epsilon^{-1}} in the O(1/\epsilon)-smooth case. Our methods consist of a recently proposed ball optimization oracle acceleration algorithm (which we refine) and a careful implementation of said oracle for the softmax function. We also prove an oracle complexity lower bound scaling as \Omega(N*\epsilon^{-2/3}), showing that our dependence on N is optimal up to polylogarithmic factors. This talk is based on joint work with Yair Carmon, Arun Jambulapati, and Aaron Sidford. Cheers, David -------------- next part -------------- An HTML attachment was scrubbed... URL: From wajc at stanford.edu Thu Aug 26 11:56:16 2021 From: wajc at stanford.edu (David Wajc) Date: Thu, 26 Aug 2021 11:56:16 -0700 Subject: [theory-seminar] Theory Lunch 08/26: Yujia Jin In-Reply-To: References: Message-ID: Gentle reminder: last theory lunch of the quarter starting in a few minutes! On Mon, 23 Aug 2021 at 13:12, David Wajc wrote: > Hi all, > > The last theory lunch of the summer quarter will take place Thursday at > noon (PDT), at this zoom room > . As > usual, we'll start with some socializing, followed by a talk starting at > 12:30pm. > Yujia will tell us about: *Thinking Inside the Ball: Near-Optimal > Minimization of the Maximal Loss* > > *Abstract:* In this work, we characterize the complexity of minimizing > the maximal of N functions for convex, Lipschitz functions f_1,? , f_N. > For non-smooth functions, existing methods require O(N*\epsilon^{-2}) > queries to a first-order oracle to compute an \epsilon-suboptimal point > and \Otil{N*\epsilon^{-1}} queries if the functions f_i are > O(1/\epsilon)-smooth. > > We develop methods with improved complexity bounds of > \Otil{N*\epsilon^{-2/3} + \epsilon^{-8/3}} in the non-smooth case and > \Otil{N*\epsilon^{-2/3} + \sqrt{N}*\epsilon^{-1}} in the > O(1/\epsilon)-smooth case. Our methods consist of a recently proposed ball > optimization oracle acceleration algorithm (which we refine) and a careful > implementation of said oracle for the softmax function. We also prove an > oracle complexity lower bound scaling as \Omega(N*\epsilon^{-2/3}), showing > that our dependence on N is optimal up to polylogarithmic factors. > > This talk is based on joint work with Yair Carmon, Arun Jambulapati, and > Aaron Sidford. > > Cheers, > David > -------------- next part -------------- An HTML attachment was scrubbed... URL: From wajc at stanford.edu Thu Aug 26 11:56:16 2021 From: wajc at stanford.edu (David Wajc) Date: Thu, 26 Aug 2021 11:56:16 -0700 Subject: [theory-seminar] Theory Lunch 08/26: Yujia Jin In-Reply-To: References: Message-ID: Gentle reminder: last theory lunch of the quarter starting in a few minutes! On Mon, 23 Aug 2021 at 13:12, David Wajc wrote: > Hi all, > > The last theory lunch of the summer quarter will take place Thursday at > noon (PDT), at this zoom room > . As > usual, we'll start with some socializing, followed by a talk starting at > 12:30pm. > Yujia will tell us about: *Thinking Inside the Ball: Near-Optimal > Minimization of the Maximal Loss* > > *Abstract:* In this work, we characterize the complexity of minimizing > the maximal of N functions for convex, Lipschitz functions f_1,? , f_N. > For non-smooth functions, existing methods require O(N*\epsilon^{-2}) > queries to a first-order oracle to compute an \epsilon-suboptimal point > and \Otil{N*\epsilon^{-1}} queries if the functions f_i are > O(1/\epsilon)-smooth. > > We develop methods with improved complexity bounds of > \Otil{N*\epsilon^{-2/3} + \epsilon^{-8/3}} in the non-smooth case and > \Otil{N*\epsilon^{-2/3} + \sqrt{N}*\epsilon^{-1}} in the > O(1/\epsilon)-smooth case. Our methods consist of a recently proposed ball > optimization oracle acceleration algorithm (which we refine) and a careful > implementation of said oracle for the softmax function. We also prove an > oracle complexity lower bound scaling as \Omega(N*\epsilon^{-2/3}), showing > that our dependence on N is optimal up to polylogarithmic factors. > > This talk is based on joint work with Yair Carmon, Arun Jambulapati, and > Aaron Sidford. > > Cheers, > David > -------------- next part -------------- An HTML attachment was scrubbed... URL: From monika.henzinger at univie.ac.at Mon Aug 30 13:25:05 2021 From: monika.henzinger at univie.ac.at (Monika Henzinger) Date: Mon, 30 Aug 2021 13:25:05 -0700 Subject: [theory-seminar] CS 369Z Fall 2021 Message-ID: Hi all, I am a visiting professor at Stanford and will teach CS 369Z this fall semester.? Here is the announcement, please come if you are interested in dynamic data structures. Best regards, Monika CS 369Z: Dynamic data structures for graphs and point sets Instructor: Monika Henzinger Fall 2021 Date: Tue/Th 9:45 ? 11:15am Location: Hewlett Teaching Center Rm. 103 Grading: Pass/Fail Description: With the increase of huge, dynamically changing data sets there is a raisingneed for dynamic data structures to represent and process them. This course will present the algorithmic techniques that have been developed for dynamic data structures for graphs and for point sets. Syllabus: 1. Computational models: * Adversarial models, average-case model, cell probe model, pointer machine model, oracle model, conditional lower bounds 2. Fundamental data structures: * Euler-Tour trees and Top Trees * Epsilon-nets, cover trees, range-search data structures 3. Hierarchy-based dynamic graph algorithms: * Connectivity and minimum spanning trees, approximate cardinality matching, vertex coloring, shortest paths 4. Random-rank based dynamic graph algorithms: * Maximal independent set, maximal matching, vertex coloring 5. Point Sets 1: Dynamic proximity search * Different variants of approximate nearest neighbor 6. Point Sets 2: Dynamic clustering algorithms: * Core sets, k-center, k-means, k-median, hierarchical clustering -------------- next part -------------- An HTML attachment was scrubbed... URL: