Adrian Markelov:
This week I have been reading papers on time series classification (see presentation for samples) and reviewing deep learning design with pytorch from previous projects I have done. I would also like to answer a question asked from our presentation review. Why PCA before the deep net. The reason we need to compress the dataset before we put it through the deep net is simply because we do not have that much data. When you have any machine learning model that has a lot of parameters that need to be learned it is essential that the amount of training data vastly exceeds the number of parameters that need to be learned. If this requirement is not satisfied that model will over fit the given training data and will fail on then validation set and thus in actual practical use. Over time we will gather a large amount of data and remove PCA when it seems appropriate. Some simple testing with cross validation will tell use when we have enough data to not overfit.
Kyle:
This week was mostly slack for me, focusing on the design review and finishing up assignments for other classes.