Exercise 10: Machine learning based components (Lab 3 prep)

Computer Exercise (Mandatory for Lab 3): Design a neural network to recognize an omnibot

You find an aptly-named notebook (Exercise_10.ipynb) on the jupyterhub servers under the latest version of frtn75-vision-sensor-and-ilc.

 

Optional Exercises:

Neural Network Playground

Spend some time with the NN playground, Links to an external site. if you have not done it before
Challenge: Find architecture and parameters so you get below 2% test loss on the spiral set in less than 1000 training epochs. Also try to minimize the number of used neurons. Hint: Note that you can vary the learning rate while the training is running.

Time series prediction for the Lorentz ODE in matlab

Download, study and run the code NNpred.m Download NNpred.m and make sure you understand it. Try changing the NN architecture (add or decrease nr of layers, change activation functions, vary  size of hidden parameters). Can you decrease the network size or training time, while maintaining similar prediction accuracy?

Time series prediction with LSTM in matlab

Download, study and run the code LSTMchickenpox.m Download LSTMchickenpox.m and make sure you understand it. How does the size of hidden state (numHiddentUnits) impact the accuracy and training times.