[Colloquium] REMINDER: 2/5 Talks at TTIC: Holden Lee, Princeton University

[Mary Marre via Colloquium]

When:     Tuesday, February 5th at *11:00 am*

Where:    TTIC, 6045 S Kenwood Avenue, 5th Floor, Room 526

Who:       Holden Lee, Princeton University


*Title:       *Provable Algorithms for Sampling and for Learning Linear
Dynamical Systems

*Abstract:* A key problem in Bayesian machine learning is to sample from a
probability distribution whose density is specified up to a partition
function, $p(x)\propto e^{-f(x)}$. Two shortcomings of current Markov Chain
Monte Carlo methods (such as Langevin Monte Carlo) which limit their
applicability are that (1) probability distributions in practice are
*multimodal*, and (2) often the distribution needs to be be updated in an
*online* fashion in response to streaming data. I prove that (1) combining
Langevin diffusion with temperature-based methods can exponentially speed
up mixing for multimodal distributions (using a new Markov chain
decomposition theorem), and (2) variance-reduced stochastic gradient
methods allow online sampling in almost-constant time per update.

Reinforcement learning has produced some of the most impressive results in
machine learning, but lacks the theoretical guarantees of control-theoretic
algorithms. I describe my work in bridging this gap in the case of *linear
dynamical systems with hidden state* by developing algorithms that can both
learn the system (like model-based RL) and provide guarantees on its
control (as in control theory).

Covers joint work with Sanjeev Arora, Rong Ge, Elad Hazan, Oren Mangoubi,
Andrej Risteski, Karan Singh, Nisheeth Vishnoi, Cyril Zhang, and Yi Zhang.


Host: Nathan Srebro <nati at ttic.edu>


Mary C. Marre
Administrative Assistant
*Toyota Technological Institute*
*6045 S. Kenwood Avenue*
*Room 517*
*Chicago, IL  60637*
*p:(773) 834-1757*
*f: (773) 357-6970*
*mmarre at ttic.edu <mmarre at ttic.edu>*


On Mon, Feb 4, 2019 at 3:43 PM Mary Marre <mmarre at ttic.edu> wrote:

> When:     Tuesday, February 5th at *11:00 am*
>
> Where:    TTIC, 6045 S Kenwood Avenue, 5th Floor, Room 526
>
> Who:       Holden Lee, Princeton University
>
>
> *Title:       *Provable Algorithms for Sampling and for Learning Linear
> Dynamical Systems
>
> *Abstract:* A key problem in Bayesian machine learning is to sample from
> a probability distribution whose density is specified up to a partition
> function, $p(x)\propto e^{-f(x)}$. Two shortcomings of current Markov Chain
> Monte Carlo methods (such as Langevin Monte Carlo) which limit their
> applicability are that (1) probability distributions in practice are
> *multimodal*, and (2) often the distribution needs to be be updated in an
> *online* fashion in response to streaming data. I prove that (1) combining
> Langevin diffusion with temperature-based methods can exponentially speed
> up mixing for multimodal distributions (using a new Markov chain
> decomposition theorem), and (2) variance-reduced stochastic gradient
> methods allow online sampling in almost-constant time per update.
>
> Reinforcement learning has produced some of the most impressive results in
> machine learning, but lacks the theoretical guarantees of control-theoretic
> algorithms. I describe my work in bridging this gap in the case of *linear
> dynamical systems with hidden state* by developing algorithms that can both
> learn the system (like model-based RL) and provide guarantees on its
> control (as in control theory).
>
> Covers joint work with Sanjeev Arora, Rong Ge, Elad Hazan, Oren Mangoubi,
> Andrej Risteski, Karan Singh, Nisheeth Vishnoi, Cyril Zhang, and Yi Zhang.
>
>
> Host: Nathan Srebro <nati at ttic.edu>
>
>
>
> Mary C. Marre
> Administrative Assistant
> *Toyota Technological Institute*
> *6045 S. Kenwood Avenue*
> *Room 517*
> *Chicago, IL  60637*
> *p:(773) 834-1757*
> *f: (773) 357-6970*
> *mmarre at ttic.edu <mmarre at ttic.edu>*
>
>
> On Tue, Jan 29, 2019 at 8:56 PM Mary Marre <mmarre at ttic.edu> wrote:
>
>> When:     Tuesday, February 5th at *11:00 am*
>>
>> Where:    TTIC, 6045 S Kenwood Avenue, 5th Floor, Room 526
>>
>> Who:       Holden Lee, Princeton University
>>
>>
>> *Title:       *Provable Algorithms for Sampling and for Learning Linear
>> Dynamical Systems
>>
>> *Abstract:* A key problem in Bayesian machine learning is to sample from
>> a probability distribution whose density is specified up to a partition
>> function, $p(x)\propto e^{-f(x)}$. Two shortcomings of current Markov Chain
>> Monte Carlo methods (such as Langevin Monte Carlo) which limit their
>> applicability are that (1) probability distributions in practice are
>> *multimodal*, and (2) often the distribution needs to be be updated in an
>> *online* fashion in response to streaming data. I prove that (1) combining
>> Langevin diffusion with temperature-based methods can exponentially speed
>> up mixing for multimodal distributions (using a new Markov chain
>> decomposition theorem), and (2) variance-reduced stochastic gradient
>> methods allow online sampling in almost-constant time per update.
>>
>> Reinforcement learning has produced some of the most impressive results
>> in machine learning, but lacks the theoretical guarantees of
>> control-theoretic algorithms. I describe my work in bridging this gap in
>> the case of *linear dynamical systems with hidden state* by developing
>> algorithms that can both learn the system (like model-based RL) and provide
>> guarantees on its control (as in control theory).
>>
>> Covers joint work with Sanjeev Arora, Rong Ge, Elad Hazan, Oren Mangoubi,
>> Andrej Risteski, Karan Singh, Nisheeth Vishnoi, Cyril Zhang, and Yi Zhang.
>>
>>
>> Host: Nathan Srebro <nati at ttic.edu>
>>
>>
>>
>> Mary C. Marre
>> Administrative Assistant
>> *Toyota Technological Institute*
>> *6045 S. Kenwood Avenue*
>> *Room 517*
>> *Chicago, IL  60637*
>> *p:(773) 834-1757*
>> *f: (773) 357-6970*
>> *mmarre at ttic.edu <mmarre at ttic.edu>*
>>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mailman.cs.uchicago.edu/pipermail/colloquium/attachments/20190205/57eb8f8e/attachment-0001.html>