Hi,

Yes, that sounds good. I will start working on UI aspects while waiting for
them to release fixes. If they release fixes while working on the UI it
should be fine. Otherwise, I will start implementing the autoencoder.

On Thu, Jun 18, 2015 at 12:04 PM, Srinath Perera <[email protected]> wrote:

> Let's work on other parts of the project, for example we can look at UI to
> configure Deep learning while waiting for them to do it. However, fix takes
> time, we have to write auto encoder.
>
> On Thu, Jun 18, 2015 at 7:26 AM, Thushan Ganegedara <[email protected]>
> wrote:
>
>> Dear all,
>>
>> I posted about poor accuracy on deeplearning4j google group and they
>> replied saying they are aware of this issue (
>> https://groups.google.com/forum/#!topic/deeplearning4j/gRQH8rA9SQI).
>> They are currently working on it, to get the tuned examples out soon.
>>
>> However, from the appearance of their response, deeplearning4j has few
>> bugs/issues with gradient calculation after they moved to nd4j.
>>
>> I can see two options right now,
>>
>> 1. Wait till the bugs are fixed and tuned examples are out
>> 2. Write autoencoder myself, adhering to the coding style they have
>> followed. So the example can be easily integrated with deeplearning4j after
>> a stable version is out.
>>
>>
>> Would highly appreciate your feedback.
>>
>> On Wed, Jun 17, 2015 at 10:27 AM, Thushan Ganegedara <[email protected]>
>> wrote:
>>
>>> Hi,
>>>
>>> I couldn't find any reports on high accuracy on any set of data. The
>>> popular datasets seem to be Iris dataset and MNIST dataset. However, their
>>> example page on git hub shows several results for different techniques and
>>> datasets (https://github.com/deeplearning4j/dl4j-0.0.3.3-examples). And
>>> as it can be seen, the accuracies mentioned are quite low. (However, the
>>> developers claim that this is because they have run the examples with less
>>> number of nodes and less iterations)
>>>
>>>
>>>
>>> On Tue, Jun 16, 2015 at 6:03 PM, Srinath Perera <[email protected]>
>>> wrote:
>>>
>>>> Are there any other datasets where dl4j suppose to do well? As long as
>>>> it does better with *some* datasets, we can go ahead with those?
>>>>
>>>> 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
>>>>>
>>>>
>>>>
>>>>
>>>> --
>>>> ============================
>>>> Blog: http://srinathsview.blogspot.com twitter:@srinath_perera
>>>> Site: http://people.apache.org/~hemapani/
>>>> Photos: http://www.flickr.com/photos/hemapani/
>>>> Phone: 0772360902
>>>>
>>>
>>>
>>>
>>> --
>>> Regards,
>>>
>>> Thushan Ganegedara
>>> School of IT
>>> University of Sydney, Australia
>>>
>>
>>
>>
>> --
>> Regards,
>>
>> Thushan Ganegedara
>> School of IT
>> University of Sydney, Australia
>>
>
>
>
> --
> ============================
> Blog: http://srinathsview.blogspot.com twitter:@srinath_perera
> Site: http://people.apache.org/~hemapani/
> Photos: http://www.flickr.com/photos/hemapani/
> Phone: 0772360902
>



-- 
Regards,

Thushan Ganegedara
School of IT
University of Sydney, Australia
_______________________________________________
Dev mailing list
[email protected]
http://wso2.org/cgi-bin/mailman/listinfo/dev

Reply via email to