Applying Jackknife and the need to reduce input layers (X,Y) by one - Regression


My ANN’s predicts numerical values as output. The output layer is of length 4 (since it predicts 4 values from data). My input is X and Y, each with a length of 1000.


To put an uncertainty on each of my predicted values from my ANN, I want to apply jackknife. That is to evaluating my model several times and deleting one random datapoint in each evaluation. Thus, I would need to change my input layers (X and Y) to a length of 999.

Is it possible to do this? If not, why not?


Some insights/ideas would be very much appreciated here… :see_no_evil:

It is not clear to me what you are aiming to. The layers shape is constant, and the reason is that each of the X values will be multiplied by a weight.

To evaluate if the net is doing well, just use a test set. Why isn’t this useful ?


Thanks for your reply.

Getting an uncertainty for each parameter is possible with K-fold validation (which I’m applying) but then I just get K predictions with a ton of computation/training time whereas with jackknife (or bootstrapping) I get multiple predictions by just using a single test, applying n evaluations with jackknife, and then appending the predictions each evaluation.


1 Like