1 minute read

What I did

  • Sunday: Trained the neural network on the skew-norm dataset and spent some time playing around with the architecture and hyperparameters.
  • Monday: I picked a model from yesterday I was happy with and used it to predict the data from the particle-averaged (real disorder) dataset. The results were dissapointing. I also spent a lot of time working on job interview preparation.
  • Tuesday: I’m investigating why the NN performed poorly predicting the other dataset. I have some theories and ideas for solutions, but I want to pull up some specific examples to illustrate why.
  • Wednesday - Sunday: Re-thought how to define MSD, both in terms of calculating it for the particle averaged structure and in terms of how I’m weighting the skew-norm averaged ones. This seems to have fixed my issues. This is now going to be published work, so I am going to be vague on details to protect my work.

What I learned

  • Sunday: A simple 2 hidden layer architecture actually works really well for signal processing, essentially picking out minute distinctions between spectra.
  • Monday: I was helping out a friend with a pandas problem where he couldn’t extract the value from a row, and was getting an ndarray or a Series object. The solution was to add .item() to the end of the .loc[" "][ ] statement.
  • Tuesday: Here’s my thoughts behind the poor generalization. My NN is predicting MSD, an aggregate statistic gerenated from a clever skew-norm averaging. However, 2 very different skew-norm distributions can produce similar or even identical MSD values, yet create wildly different looking structures. I might be able to improve performance by actually having the NN predict the moments of the skew-norm distribution. This would be equivilient to predicting the full radial dsitribution frunction \(g(r)\). One other thing I’ve noticed, however, is that the predictions are consistently low. It seems that the shapes of the particle averaged spectra and pretty different from the skew-norm averaged ones, and might require some scaling function to better match. Interestingly, the gaussian averaged ones matched quite well. Not sure why the skew-norm are so different.

What I will do next

  • Write my abstract for my thesis defense. Improve NN to predict multiple outputs.

Categories:

Updated: