The results in the paper from which this algorithm was implemented are 
encouraging: 
http://www.research.rutgers.edu/~lihong/pub/Zinkevich11Parallelized.pdf

The proof is a bit beyond me so I cannot vouch for the theory.  I'm excited 
to test this on some non - trivial problems to see how it fares.

_Chris

On Monday, June 9, 2014 12:19:39 PM UTC-4, John Myles White wrote:
>
> My question is about the theory behind your algorithm. My understanding is 
> that no parallel SGD implementation (except one that trivially runs on the 
> same data) will produce correct results in general. Is that not true?
>
>  -- John
>
> On Jun 9, 2014, at 9:07 AM, Christopher Fusting <[email protected] 
> <javascript:>> wrote:
>
> John, 
>
> There has been no rigorous testing yet.  My primary concern in the 
> averaging algorithm is process latency, completion time, and faults.  Do 
> you have specifics you would like to share?  
>
> _Chris
>
>
> On Mon, Jun 9, 2014 at 11:24 AM, John Myles White <[email protected] 
> <javascript:>> wrote:
>
>> Very cool, Chris.
>>
>> I’ve done a lot of work on SGD in Julia, so I’m glad to see more.
>>
>> Regarding the averaging technique you’re using, have you done much 
>> testing to see how well it works? My sense is that the algorithm you’re 
>> using is a little brittle, but perhaps I’ve misunderstood it.
>>
>>  — John
>>
>> On Jun 8, 2014, at 11:36 AM, Christopher Fusting <[email protected] 
>> <javascript:>> wrote:
>>
>> Hi everyone.  I've been playing around with Julia for awhile now and have 
>> implemented Parallel Stochastic Gradient Descent.  This is my first Julia 
>> project (and attempt at implementing this algorithm) so its not perfect, 
>> but I think I have a good start and wanted to share it: 
>> https://github.com/cfusting/PSGD.jl.  I welcome any feedback.
>>
>> Eventually I'd like to integrate the package with DataFrames and do a 
>> little optimization, especially on the algorithm that partition the data.
>>
>> _Chris
>>
>>
>>
>
>
> -- 
> Christopher W. Fusting
> *Software Developer / Analyst*
>
> @cfusting
> 828-772-0012
>  
>
>

Reply via email to