Hi,

Ok, then I will continue with the test accuracy in mind.

On Fri, Jun 19, 2015 at 2:01 PM, Nirmal Fernando <[email protected]> wrote:

> Hi Thushan,
>
> Yes, so currently we take a training data fraction as an input from the
> user (check the wizard) which splits the dataset to training and test. And
> for other algorithms too, Spark generates an accuracy measurement using the
> test data predictions vs actual.
>
> test *accuracy* sounds good as a measurement.
>
> On Fri, Jun 19, 2015 at 9:19 AM, Thushan Ganegedara <[email protected]>
> wrote:
>
>> Hi all,
>>
>> Thank you very much for the feedback
>>
>> One small thing, is there a some sort of accuracy measurement that we can
>> show for deep networks?
>>
>> Yes, there is. Usually the accuracy of the deep network is shown with a
>> validation dataset and a test set (i.e. validation error and test error).
>> In other words, after training the network, we run an independent test set
>> and see how accurate the algorithm is (# of correct results/ # of total
>> results)
>>
>> It seems the test accuracy would fit for the model comparison
>>
>> On Fri, Jun 19, 2015 at 1:01 PM, Nirmal Fernando <[email protected]> wrote:
>>
>>> Hi Thushan,
>>>
>>> Looks good for me too. One small thing, is there a some sort of accuracy
>>> measurement that we can show for deep networks? This is required for the
>>> model comparison page;
>>> https://docs.wso2.com/display/ML100/Model+Comparison
>>>
>>> On Fri, Jun 19, 2015 at 7:27 AM, Thushan Ganegedara <[email protected]>
>>> wrote:
>>>
>>>> Hi,
>>>>
>>>> as he configure, can we visualize the deep network ( use some d3 graph
>>>> library).
>>>> I don't think I understood. Are you asking if we could show the user a
>>>> diagram of the network he has specified (like the attached image)?
>>>>
>>>
>>>> Also, after he has trained, can we let him visualize the output of
>>>> intermediate layers ( this is a advanced feature, so optional).
>>>> Intermediate levels are quite tricky to visualize. Two major issues
>>>> with visualizing intermediate layers are,
>>>> 1. Intermediate layers are not linear transformations of the input so
>>>> the visualized filters of intermediate layers doesn't make much sense (I
>>>> tried it, what I saw was completely random pixels)
>>>>
>>>> 2. It is possible to overcome the above issue with a technique called
>>>> activation maximization (we are solving the optimization problem z =
>>>> sigmoid(x.W) by keeping W constant and changing x to get the maximum z).
>>>> However, this is highly non-convex. So it easily get stuck in local maxima.
>>>> Also, this is to costly for large dimensional data (it will work for MNIST
>>>> though).
>>>>
>>>> Thank you
>>>>
>>>> On Fri, Jun 19, 2015 at 11:50 AM, Srinath Perera <[email protected]>
>>>> wrote:
>>>>
>>>>> Hi Tushan,
>>>>>
>>>>> OK high level. However, as he configure, can we visualize the deep
>>>>> network ( use some d3 graph library).
>>>>>
>>>>> Also, after he has trained, can we let him visualize the output of
>>>>> intermediate layers ( this is a advanced feature, so optional).
>>>>>
>>>>> --Srinath
>>>>>
>>>>> On Fri, Jun 19, 2015 at 6:56 AM, Thushan Ganegedara <[email protected]>
>>>>> wrote:
>>>>>
>>>>>> Dear all,
>>>>>>
>>>>>> Please find the proposed UI changes for the Deep Network Integration
>>>>>> attached herewith.
>>>>>>
>>>>>> Feedback would be highly appreciated.
>>>>>>
>>>>>> --
>>>>>> 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
>>>>
>>>
>>>
>>>
>>> --
>>>
>>> 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
>>
>
>
>
> --
>
> 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

Reply via email to