LeNet CNN Architecture Quantitative Metrics

[Adamu Aminu]

Hello @all Opendtect Users,

I did seismic facies classification using the deep learning CNN architecture. I am interested in getting additional quantitative metrics (e.g., confusion matrix, precision, recall, F1-score, or class-wise accuracy) in addition to the validation accuracy and validation loss. Is there provision for any of such metrics and how can i get it? 

Thank you for your prompt response. 

[Mark Crawford]

Hi Adamu,

In our original Neural Network plugin, which is started from the Neural Networks button in the ML Control Center, we produce a statistical report for MLP classifications from which you can create a confusion plot.

Kind regards,

Mark

Mark Crawford

Support Manager

-----------------------------------------------------------------------------------------------------------------------   dGB Earth Sciences
Phone: +31 53 4315155 E-mail: mark.c...@dgbes.com Internet: dgbes.com

----------------------------------------------------------------------------------------------------------------------- 

--
You received this message because you are subscribed to the Google Groups "OpendTect Users" group.
To unsubscribe from this group and stop receiving emails from it, send an email to users+un...@opendtect.org.
To view this discussion visit https://groups.google.com/a/opendtect.org/d/msgid/users/e2ffce61-92d1-4b9a-b096-5a100805d8edn%40opendtect.org.

[Adamu Aminu]

Hi Mark,

Thank you for the response, if i get you right the confusion plot in the Neural Network button is for MultiLayer Perceptron, can it also serve for  LeNet CNN architecture?

My regards to Paul de Groot, all of you at  dGB Earth Science and @all Opendtect Users.

To view this discussion visit https://groups.google.com/a/opendtect.org/d/msgid/users/CAKpeyVPYLLqTaA0w1s7JHQggR6j0wLyZ94ktOy5NQPcnScycUA%40mail.gmail.com.

[Uwe Strecker]

Dear Adamu,

Thank you for your question. The quantitative metrics that can be reported depend on whether the network is supervised or unsupervised. For a CNN or MLP trained in supervised mode, where labelled training data are available, OpendTect supports evaluation using validation accuracy and validation loss during training. In addition, confusion plots are supported for MLPs. Because supervised models have known target labels, they allow the computation of true positives, false positives, true negatives, and false negatives, from which metrics such as precision, recall, F1-score, and class-wise accuracy can be derived. These statistics are available through the trained network information and confusion plot.

For unsupervised methods such as UVQ (Unsupervised Vector Quantizer primarily used for seismic facies analysis via Waveform Segmentation), these metrics are not applicable because the algorithm does not use predefined class labels. Instead, UVQ groups similar seismic responses into clusters based on their similarity, and the interpretation of those clusters is performed after training. Since there is no ground truth during training, a confusion matrix or related classification metrics cannot be calculated. Consequently, there is also no direct quantitative comparison between an unsupervised UVQ result and a supervised CNN/MLP result using confusion-matrix-based statistics. For unsupervised networks the match attribute serves as metric. As you know, the match is the normalised Euclidean distance between the input vector and the vector representing the winning class. 1 means the vectors are identical, 0 means they are completely dissimilar.

If your goal is to report precision, recall, F1-score, or class-wise accuracy, a supervised workflow (such as an MLP) is the appropriate choice. For CNN models, if you require these additional metrics beyond validation accuracy and loss, they would need to be computed externally using the model predictions and the corresponding ground-truth labels.

--

You received this message because you are subscribed to the Google Groups "OpendTect Users" group.
To unsubscribe from this group and stop receiving emails from it, send an email to users+un...@opendtect.org.
To view this discussion visit https://groups.google.com/a/opendtect.org/d/msgid/users/e2ffce61-92d1-4b9a-b096-5a100805d8edn%40opendtect.org.

--

 Uwe Strecker, Ph.D.

Geoscience Manager

--------------------------------------------------------------------------------------------------------------------   dGB Earth Sciences
Phone: +31 53 4315155 E-mail: uwe.st...@dgbes.com Internet: dgbes.com

----------------------------------------------------------------------------------------------------------------------

[Adamu Aminu]

Thank you so much for your time and the comprehensive and satisfied answer rendered. 

Kind Regards.

To view this discussion visit https://groups.google.com/a/opendtect.org/d/msgid/users/CABeQnOag17uYsc5h41Lrg-euMtj4bMOmYnaPyjoap7%3DbvVdjjg%40mail.gmail.com.