Data Mining expert
Budget: $10 – $30 USD
1. Draw a network architecture that includes few fully connected layers and is able to perform classification on the Fashion MNIST dataset: https://github.com/zalandoresearch/fashion-mnist.
Define the network to your liking - it can be as simple or as complex as you can make it. The network will be considered correct if all dimensions involved are fully specified.
2. Train the network in Colab (you need to enable the GPU runtime) and plot the average loss vs SGD iteration as well as its precision vs recall curve. Clearly state and show your hyperparameter optimization strategy.
3. Randomly shuffle (permute) the pixels of the input images such that the pixel $(i,j)$ is randomly replaced with another pixel $(k,l)$ and repeat the training exercise, plotting the average loss vs SGD iteration number as well as the precision vs recall curve.
4. Randomly shuffle (permute) the labels such that the label $y_j$ is replaced with another $y_k$ (obviously though within the range 0-9) and repeat the training exercise, plotting plotting the average loss vs SGD iteration number as well as the precision vs recall curve.
5. Compare your findings against the findings of [this important paper](https://arxiv.org/pdf/1611.03530.pdf)
Define the network to your liking - it can be as simple or as complex as you can make it. The network will be considered correct if all dimensions involved are fully specified.
2. Train the network in Colab (you need to enable the GPU runtime) and plot the average loss vs SGD iteration as well as its precision vs recall curve. Clearly state and show your hyperparameter optimization strategy.
3. Randomly shuffle (permute) the pixels of the input images such that the pixel $(i,j)$ is randomly replaced with another pixel $(k,l)$ and repeat the training exercise, plotting the average loss vs SGD iteration number as well as the precision vs recall curve.
4. Randomly shuffle (permute) the labels such that the label $y_j$ is replaced with another $y_k$ (obviously though within the range 0-9) and repeat the training exercise, plotting plotting the average loss vs SGD iteration number as well as the precision vs recall curve.
5. Compare your findings against the findings of [this important paper](https://arxiv.org/pdf/1611.03530.pdf)