Hi, Thank you for your feedback.
As for the moment I haven't tried CNN. However, we do not need CNNs for the MNIST dataset. With a 3-layer SAE (Stacked AE) we can easily reach an accuracy of 98%+ (Without any special regularization). As I've tested this in python with a hand-written SAE Also, CNN creates quite an ovehead with many sub-steps (convolution, subsampling, pooling) as far as I know about it. CNNs do quite a good job for high-resolution images, but in this case it will create a substantial overhead. However, I will try it and see. Thank you for your suggestion On Tue, Jun 16, 2015 at 4:09 PM, CD Athuraliya <[email protected]> wrote: > Hi Thushan, > > Assuming your are running examples, did you try CNN for MNIST data? What > is the accuracy? > > You can use DL4J Gitter chat room [1] for quick responses from devs. > > [1] https://gitter.im/deeplearning4j/deeplearning4j > > Regards, > CD > > On Tue, Jun 16, 2015 at 9:13 AM, Thushan Ganegedara <[email protected]> > wrote: > >> Yes, there are few use list (Git hub and google group). I will inquire >> about this in user lists. >> >> Thank you >> >> >> On Tue, Jun 16, 2015 at 12:34 PM, Nirmal Fernando <[email protected]> >> wrote: >> >>> Thanks Thushan for the update. >>> >>> In addition to you digging into the code, can you also inquire on the >>> poor performance from the DL4J user list (if there's one exist)? >>> >>> On Tue, Jun 16, 2015 at 5:27 AM, Thushan Ganegedara <[email protected]> >>> wrote: >>> >>>> Dear all, >>>> >>>> Please find the update regarding DL4J testing >>>> >>>> *Poor Accuracy* >>>> I have been testing DL4J extensively with *MNIST and Iris* datasets >>>> (Small and Full). However, I was unable to get a reasonable accuracy with >>>> DL4J for the aforementioned datasets. The F1-score was around 0.02, which >>>> is very low. >>>> >>>> I tried with different settings mainly for the following attributes >>>> >>>> Weight initialization >>>> Gradient Descent >>>> Iterations >>>> Type of units: Autoencoder/RBM >>>> >>>> >>>> But none of the settings gave a reasonable accuracy. Furthermore, the >>>> predicted values for the test data usually *belong to 1 or 2 classes *(e.g. >>>> when trained on MNIST dataset, the program predict 0 and 1 only, though >>>> there are 10 possible classes) >>>> >>>> Also there are many reports of *poor accuracy of DL4J.* The best >>>> accuracy I could find reported was around 0.5 F1 score for MNIST, which is >>>> still very low. (e.g. MNIST can easily reach 0.9+ accuracy for even a >>>> basic deep network) >>>> >>>> I'm currently trying to delve in to the code for DL4J and figure out >>>> how the learning is done. I'm assuming there are some faults in the >>>> learning process which causes the algorithm to learn poorly. >>>> >>>> Thank you >>>> >>>> -- >>>> Regards, >>>> >>>> Thushan Ganegedara >>>> School of IT >>>> University of Sydney, Australia >>>> >>> >>> >>> >>> -- >>> >>> Thanks & regards, >>> Nirmal >>> >>> Associate Technical Lead - Data Technologies Team, WSO2 Inc. >>> Mobile: +94715779733 >>> Blog: http://nirmalfdo.blogspot.com/ >>> >>> >>> >> >> >> -- >> Regards, >> >> Thushan Ganegedara >> School of IT >> University of Sydney, Australia >> >> _______________________________________________ >> Dev mailing list >> [email protected] >> http://wso2.org/cgi-bin/mailman/listinfo/dev >> >> > > > -- > *CD Athuraliya* > Software Engineer > WSO2, Inc. > lean . enterprise . middleware > Mobile: +94 716288847 <94716288847> > LinkedIn <http://lk.linkedin.com/in/cdathuraliya> | Twitter > <https://twitter.com/cdathuraliya> | Blog > <http://cdathuraliya.tumblr.com/> > -- Regards, Thushan Ganegedara School of IT University of Sydney, Australia
_______________________________________________ Dev mailing list [email protected] http://wso2.org/cgi-bin/mailman/listinfo/dev
