Network Probing methods

[voetberg]

After discussion in the workshop - it would be advantageous to be able to monitor the message passing after each iteration

• During training -> include a static decoder that looks at the messages after each iteration (basically, an advanced loss curve).
• After training -> Stick this static decoder in after each step of linear planes, see where the most important part of the network is (similar to a relevance walk)

[voetberg]

Progress update:

• Completed the probes and writing the network-stepping process
• Need to decide the "percent explained" metric (time to do some reading on information theory....again)
• Need to decide the loss metric

[voetberg]

Yayyyy it works

[voetberg]

Meeting questions:

• I'm missing points for the entropy plot, do I have the backwards message passing step for the nexus plane?
• Can I document the importance of message passing steps without the probe? The probe implies that they're pointless, but this contradicts the previous optimization study.
• I need to verify the probe doesn't influence any prediction or bias the output (eg - if I apply the probe to the output, it shouldn't change the output)

[voetberg]

I was missing the backwards pass of the message passing - here it is.

Curious as to why the results changed so dramatically. I did have some shape problems I didn't have last week, I should verify the load method is picking up the network properly.

[voetberg]

Okay, this one makes much more sense.

Something to note, the linear probes are untrained, so the loss doesn't converge to 0. I'm going to reread the paper this concept is based off of and verify that this makes sense.

[voetberg]

Corrected probe shapes, it all makes sense now. Training the probes itself is a little more debatable.

[voetberg]

• Drawing some inspiration from these two papers - https://arxiv.org/pdf/1711.11279.pdf and https://arxiv.org/pdf/2107.11468.pdf ; If we can combine some of Victor's work into physics informed concepts (eg, have an output that is related to the parameters of the particle itself), we can train a probe to target that concept and see how well its included in a network. The question is how best to account for those multiple message passing steps.

[voetberg]

Question to investigate:

• How to tracks form -> Make binary labels for track-like or not track-like, impliment a simple binary loss
• where it disentangles HIPs and MIPs, -> loss would only factor in hip/mip, how heavily those are entangled at each stage

[voetberg]

While the wc cluster is down, I'm testing this with a Single data point. Loss here is binary cross entropy loss for the probe output for a point in the network. The three different planes are averaged. (Would it make more sense to split them up? Maybe).