Hello, Artyom Lyan The conversations bettwen @Stephentu and @rcurtin in this issue may be helpful. https://github.com/mlpack/mlpack/issues/370
After read the papers again, I do think the algorithm may wander around stationary point, which may be related to some parameters or initialization. Best regards, Kaiqiang On Mon, Mar 26, 2018 at 11:06 PM, <[email protected]> wrote: > Send mlpack mailing list submissions to > [email protected] > > To subscribe or unsubscribe via the World Wide Web, visit > http://knife.lugatgt.org/cgi-bin/mailman/listinfo/mlpack > or, via email, send a message with subject or body 'help' to > [email protected] > > You can reach the person managing the list at > [email protected] > > When replying, please edit your Subject line so it is more specific > than "Re: Contents of mlpack digest..." > > > Today's Topics: > > 1. Re: Fix MVU+LRSDP in GSoC 2018 (kaiqiang Xu) > 2. Regarding Proposal for Reinforcement Learning (Amit Panghal) > 3. GSOC (aditya mitkari) > 4. gsoc proposal (Артём Лян) > 5. Paper Describing Saddle Points in LRSDP (Abhijeet Krishnan) > > > ---------------------------------------------------------------------- > > Message: 1 > Date: Mon, 26 Mar 2018 01:28:18 +0800 > From: kaiqiang Xu <[email protected]> > To: mlpack_maillist <[email protected]> > Subject: Re: [mlpack] Fix MVU+LRSDP in GSoC 2018 > Message-ID: > <CABuaWqzMFc7h_h-44d99dkigEa+7fT01geR7Lp392sNKKwyhNQ@mail. > gmail.com> > Content-Type: text/plain; charset="utf-8" > > Hello, > > After I went through all the issues about MVU and LR-SDP, in particular, > #370 <https://github.com/mlpack/mlpack/issues/370>, I find that LR-SDP is > really a tough guy. It is necessary to summarise experiences of @rcurtin > and @stephentu. > I will dig into it, read source code and update my proposal. > > Best regards, > Kaiqiang > > On Sat, Mar 24, 2018 at 7:05 AM, kaiqiang Xu <[email protected]> > wrote: > > > Hi, Ryan > > > > Sorry to borther you directly by mistake in last email. > > I modified proposal and emphasize approaches. Now it has been submited to > > GSoC. Can you check it and give me some feedback if available? > > > > Best regards, > > > > Kaiqiang > > > > > > 2018-03-23 22:41 GMT+08:00 Ryan Curtin <[email protected]>: > > > >> On Fri, Mar 23, 2018 at 04:21:36PM +0800, kaiqiang Xu wrote: > >> > I have read the papers recommended under ideas-page, and been > >> fascinated by > >> > their formulations, variants(e.g. MFNN) and implementions. > >> > > >> > Firstly, MVU/MFNU is a powerful method to reduce high dimensional data > >> > which can be viewed as a more general PCA version. The paper mentions > >> that > >> > MVU/MFNU need to deal with all-nearst neighbor computation and > >> > optimization. Especially, a technique based on dual-tree and L-BFGS to > >> > solve the non-convex formulation of MVU allows MVU more scable. > >> > > >> > Secondly, SDPs is a definination to a class of optimization problems, > >> and > >> > interior point method applying to it is fast and converged. But > >> scalability > >> > is an issue. So taking advantage of low rank property of matrix, the > >> > low-rank reformulation of SDPs can be solve via Burer-Monoteiro > method. > >> > Especially, in some conditions, the Burer-Monoteiro can get global > >> optimal. > >> > > >> > Is there any error with respect to above summary? I'm glad to hear > your > >> > advice. > >> > > >> > Here I formulate some ideas: > >> > > >> > 1. > >> > > >> > Guarantee the correctness of MVU, which optimized by convex > >> optimization > >> > techniques. > >> > 2. > >> > > >> > Check the correctness of implementation of LRSDP. Design several > >> special > >> > problem, such as the Lovasz theta SDP, the maximum cut SDP > >> relaxation and > >> > etc. Solving them by LRSDP and the spectral bundle method of > >> Helmberg as > >> > well as dual-scaling interior-point method of Benson mentioned in > >> paper *A > >> > Nonlinear Programming Algorithm for Solving Semidefinite Programs > via > >> > Low-rank Factorization* . > >> > 3. > >> > > >> > Check the convergence of LRSDP under conditions mentioned in *The > >> > non-convex Burer–Monteiro approach works on smooth semidefinite > >> programs* > >> > . In paper it should go convergence. Some special case should be > >> created to > >> > test this idea. > >> > 4. > >> > > >> > Concatenate MVU and LRSDP, small dataset should be prepared to > test. > >> > Disable parameter auto-tunning, then check the convergence and > >> compare its > >> > result with idea(1) or other tools. If there is any problem > >> happended, dig > >> > into the optimization section. Hand computation is needed. My > >> intrests in > >> > optimization makes me never afraid of computation. > >> > 5. > >> > > >> > if idea(4) is convinced, a big dataset should be employed to MVU > and > >> > LRSDP. Some extreme case may happened such as traping into local > >> minima. > >> > Here the paper applying MVU to time speach dataset may help me out. > >> > > >> > How do you think of the ideas above? Can you give me advice? > Obviously I > >> > have great passion to solve it. Later hours I will submit my draft of > >> > proposal to GSoC, may you can give me some feedback. > >> > > >> > The paper *Local Minima and Convergence in Low-Rank Semidefinite > >> > Programming* is very hard to understand. I will go through it with my > >> > professor occupied optimization. I believe I can come up with new > ideas > >> and > >> > improve my plan. > >> > >> Hi Kaiqiang, > >> > >> It sounds like you have formulated a good plan. I would suggest making > >> your approach clear in your proposal, especially with what you will do > >> if MVU+LRSDP does not converge (since I do not expect it to converge). > >> > >> Thanks, > >> > >> Ryan > >> > >> -- > >> Ryan Curtin | "Wha' happened?" > >> [email protected] | - Mike LaFontaine > >> > > > > > -------------- next part -------------- > An HTML attachment was scrubbed... > URL: <http://knife.lugatgt.org/pipermail/mlpack/attachments/ > 20180326/9eb3c115/attachment-0001.html> > > ------------------------------ > > Message: 2 > Date: Sun, 25 Mar 2018 16:38:38 -0400 > From: Amit Panghal <[email protected]> > To: [email protected] > Subject: [mlpack] Regarding Proposal for Reinforcement Learning > Message-ID: > <CABGV1BGxF4EB+ikJM9MmmAQ5uQhaZjN8g56- > [email protected]> > Content-Type: text/plain; charset="utf-8" > > Hi Mentors, > > I am 1st year graduate student in computer science at New York University. > I had spent significant amount of time going through RL papers to draft a > proposal for RL algorithm to play Atari games. I have a rough understanding > of code structure in ml-pack. I have a draft proposal ready for > implementing double DQN. I am bit confused regarding, do I need to > implement the agent which using gym api's to interact with environment and > uses the current framework( > https://github.com/mlpack/mlpack/blob/master/src/mlpack/ > methods/reinforcement_learning/q_learning.hpp) > Or implement double dqn , PPO algorithms , Persistent Advantage Learning > DQN. Isnt double DQN and DQN, already implemented? > > regards, > Amit Panghal, > Courant Insitiute of Mathematical Sciences, > New York University > -------------- next part -------------- > An HTML attachment was scrubbed... > URL: <http://knife.lugatgt.org/pipermail/mlpack/attachments/ > 20180325/db0e0d59/attachment-0001.html> > > ------------------------------ > > Message: 3 > Date: 25 Mar 2018 20:40:37 -0000 > From: "aditya mitkari" <[email protected]> > To: <[email protected]> > Subject: [mlpack] GSOC > Message-ID: <[email protected]> > Content-Type: text/plain; charset="utf-8" > > Can the PCA API code be considered for parallelization using openMP? > -------------- next part -------------- > An HTML attachment was scrubbed... > URL: <http://knife.lugatgt.org/pipermail/mlpack/attachments/ > 20180325/464ad203/attachment-0001.html> > > ------------------------------ > > Message: 4 > Date: Mon, 26 Mar 2018 10:52:05 +0300 > From: Артём Лян <[email protected]> > To: mlpack <[email protected]> > Subject: [mlpack] gsoc proposal > Message-ID: <[email protected]> > Content-Type: text/plain; charset="utf-8" > > Hello mlpack mentors. > My name is Lyan Artyom. > Could you please review my proposal, that uploaded as draft at > summerofcode.withgoogle.com > Thanks in advance. > > > -- > All the best, > Artyom Lyan > -------------- next part -------------- > An HTML attachment was scrubbed... > URL: <http://knife.lugatgt.org/pipermail/mlpack/attachments/ > 20180326/bfda1d1f/attachment-0001.html> > > ------------------------------ > > Message: 5 > Date: Mon, 26 Mar 2018 11:06:19 -0400 > From: Abhijeet Krishnan <[email protected]> > To: "[email protected]" <[email protected]> > Subject: [mlpack] Paper Describing Saddle Points in LRSDP > Message-ID: <[email protected]> > Content-Type: text/plain; charset="utf-8" > > Hi Ryan, > > You had mentioned a tech note/paper by Sam Burer describing saddle points > in LRSDP. I sent him an email regarding the same, and he believed this > paper was being referred to. Is it the right one? > > Regards, > Abhijeet Krishnan > > -------------- next part -------------- > An HTML attachment was scrubbed... > URL: <http://knife.lugatgt.org/pipermail/mlpack/attachments/ > 20180326/584c614e/attachment.html> > > ------------------------------ > > Subject: Digest Footer > > _______________________________________________ > mlpack mailing list > [email protected] > http://knife.lugatgt.org/cgi-bin/mailman/listinfo/mlpack > > ------------------------------ > > End of mlpack Digest, Vol 49, Issue 56 > ************************************** >
_______________________________________________ mlpack mailing list [email protected] http://knife.lugatgt.org/cgi-bin/mailman/listinfo/mlpack
