This page is a scribble pad!

• NN on HOG
  • Baseline: 1 hidden layer with 10 nodes gives 29.83% BER
  • Note: score depends on split! But for the same split we make the following observations:
  • When hidden node is big (e.g. 1000) results are not significantly better than with few (e.g. 10-20) hidden nodes
  • Additionally, it get much slower!
  • However, results are slightly better with two hidden layers with 100 nodes. (27.86%)
  • But more layers doesn't seem to improve things further.
• NN on overFEAT
  • Baseline: 1 hidden layer with 10 nodes gives 13.22% BER
  • Due to much higher dimensionality it it much slower
  • 1 hidden layer with 100 nodes gives 10.02%
  • 2 hidden layers with 100 nodes each gives 12.54%
  • 1 hidden layer with 200 nodes gives 10.08% and is 1.5x slower than 1 h.l. w/ 100 nodes