Is there any benchmarking code for JOONE? For a university project I need to create some benchmarking code for neural networks.
To see effects on performance of different things. I am a decently advanced Java programmer, thought not as much on the maths, I could create some, and would gladly contribute them to the project if helpful.
I just finished benchmarking JOONE, Neuroph and Encog for a university project. You can see the results here.
Thanks, just a few comments after glancing at your article and having glanced at joone before. Joone's multithreaded implementation will not increase training performance. It allows each layer to update in its own thread and, while this is nice, it adds enough overhead to slow down training. It is nice, though, if some layers are connected to some external device and can update on their own. There is little to gain from updating a feed forward network in multiple threads.
If you use backprop in batch mode, then you can split the data up in chunks and in a multithreaded environment hand out chunks to each thread. At the beginning you need to dish out the chunks and at the end you need to amalgamate the results into a single update step, but apart from this overhead, the performance gain can be pretty good. I suppose this is what encog does.
Log in to post a comment.