Weekly report 3

Weekly report 3

Hours used: 10

I started questioning my previous forward propagation implementation, and simulated the matrix multiplications with pen-on-paper to be sure. I ended up updating the functionality a bit, and creating the final network class and a usage example. I also implemented a naive possibility to use batches, but it can and should probably be made in a more efficient way in the future.

I did all this to procrastinate implementing backpropagation. I read a good deal about the subject and made some notes, but actually coding it using matrices seems very unclear to me. I also struggled with storing the activations during forward propagation. Studying backpropagation involved a suprising amount of maths, and tedious derivations.

This site was especially useful with backpropagation. More often than not, the notations were quite hard to understand, like here and here. Opposite of my past experience with studying maths, wikipedia turned out to be a clear source for studying backpropagation.

Hopefully I'll finish backpropagation next week.