Challenge

• Try different values for k (passed to the KNeighborsClassifier constructor)- What can you achieve for MNIST and CIFAR-10?
• Visualize some of the errors (image, ground truth, predicted)

Bonus:

• Try different metrics (passed to the KNeighborsClassifier constructor)
• Try extracting features by decomposing the images with PCA (sklearn

Results by using KNN and PCA, and data reduction by a factor of 1/6:

• MNIST results : 0.9517%
• CIFAR10 results : 0.2889%

Challenge

• Logistic regression in 2D, reaching 93% accuracy.

Bonus:

• Logistic regression in 3D, reaching 91.6% accuracy.
• Logistic regression on MNIST, reaching 91.4% accuracy with PCA=48.

Challenge

• Get familiar with PyTorch
• Create a DL and try it on CIFAR10.

Results by using a 256-four-hiddenlayers net, heavy-dropout, large capacity:

• 57.92% accuracy on CIFAR test data.
• 76.02% accuracy on CIFAR with small CNN and data-transformations.

Challenge

• Try to solve CIFAR10 with your own numpy (MLP) neural network, i.e. implement backprop, gradient descent, batch.

Results by using a single-layer, which is "handmade" in numpy:

• 59.04% accuracy on CIFAR test data (not fine-tuned).

Challenge

• The exercise is clearly stated in the slides

• 62.25% accuracy on CIFAR test data (not fine-tuned).

Challenge

• Create a CNN and try it on CIFAR10.

• 89.47% accuracy on CIFAR with large CNN and data-transformations.

Challenge

• Create an autoencoder for celebA
• Create a VAE for mnist

• I somewhat skipped the exercise, and went for VAE on celebA.
• It seems very difficult to find the correct set of hyperparameters, which is needed to train a VAE.

Challenge

• Train an RNN to classify spam emails
• Trian an RNN to sort arrays

• I managed to do the first exercise, first, by using chars, secondly, by using words. Neither, does better than randomly guess i.e. 85% correct.