神经网络

Eumendies

测井曲线划分岩相的时候，神经网络中probability代表啥呢

Joseph

Help:

Classification (background)

Classification is used to compute discrete data and both unsupervised and supervised training can be used.

Neural Networks are only capable of dealing with continuous data. For this reason, when doing classification the algorithm will split the problem into multiple parts.

Instead of estimating the class that to be assigned to each point, it will estimate the probability of the input to belong to each class. The class with the highest probability will be assigned to that point.

A typical example of supervised classification would be to create a facies log for un-cored wells. Interpreted facies logs for the cored wells would be used for supervision to guide the classification. For instance, if you have three cored wells and three facies interpretations, then one approach could be to use two of them for supervised learning, while the third one could be compared to the estimated facies log. Another hint is to only interpret the key facies intervals (i.e. those facies intervals which you are most certain of) and use those key intervals as input for the supervised learning.

Probability threshold (background)

When using neural net classification, points can be assigned a class even though the maximum probability for the class being correct is still quite low, i.e. the probability for coarse sand occurring is only 30% but it is still the 'most probable' facies.

This can be avoided by increasing the probability threshold, i.e. to 50%. Then points where the most probable facies is less than 50% will be left undefined (gray points in the figure below).

yzli007