/deep-learning-course

Deep Learning, Udacity ud730

Assignments solutions. Note: I use sanitized validation and test set (as in Problem 5 of the first assignment). Scores:

  1. notmnist
  • LogisticRegression and OneVsAllClassifier (sklearn): test 88.0%, valid 80.0%
  1. fullyconnected
  • Multinomial classifier trained with gradient descent on partial data (10000 samples), 801 steps: test 80.7%, valid 72.9%
  • Multinomial classifier with SGD, 5001 steps: test 85.6%, valid 77.7%
  • Neural network 784 x 1024[RELU] x 10, SGD, 5001 steps: test 88.4%, valid 80.8%
  1. regularization
  • Multinomial classifier, SGD, L2 regularization, 5001 steps: test 88.4%, valid 80.3%
  • NN 784 x 1024[RELU] x 10, SGD, L2 regularization, 5001 steps: test 93.6%, valid 86.6%
  1. convolutions
  • Default covnet with 2 convs, 1 fully conected layer: 88.9% on test
  • Default covnet but with max_pool instead of simple striding: 89.6% on test