Exploiting NN parameters to construct the prediction intervals using ICP

[Waqar9871]

We have stored the NN parameters (weights and biases) corresponding to each training epoch during the model development. Upon the model training is stopped, we have the whole parameter space recorded - the size of the parameters space is dependent on the training epochs.
We used the stored parameters space hybridizing with inductive conformal prediction to construct the prediction intervals on the test dataset and call the developed techniquue Storage of Weights And Retreival Method (SWARM). We also compare the width of prediction intervals on the test dataset made by SWARM and the traditional ICP method. We see the comparable results and in some cases, SWARM based prediction intervals were tighter attributing to exploiting the local information available that adapts the prediction intervals.
For new point we want to make prediction intervals, we can use the same concept, i.e., p-value and con-conformity score. Then, use the stored parameter to estimate the prediction intervals.
More details about the method are available in paper: https://link.springer.com/article/10.1007/s41060-024-00595-w
I am happy to further collaborate to make this available to the community given your interest!

[hongxin001]

Thank you for sharing the great work. Did you open source the code implementation of this method?