Lesson 14 in-class

Part 2 (2017)
21

In the part 1 course, it was recommended RNN-GRU for Time Series compared to LSTM. Is this still true from a general Time Series analysis perspective.

22

Just to get some intuition here – is the assumption that if “time until next Promotion” is very low, sales will be depressed – because everyone is just waiting for the next upcoming sale?

23

@renjithmadhavan RNN is recommended for time series where the absolute day/week/date doesn’t matter much (like the pediatric ICU data mentioned earlier), but embeddings are recommended for structured data where day/week/date matters (like sales data)

4 Likes
24

@resdntalien that is a reasonable hypothesis

1 Like
25

Is there any issue in fitting the preprocessors on the full train and validation data? Should they be fit only on the training set?

26

For the normalization, is it possible to use another function (than log, sigmoid) ?

27

When we log transform our target variable, we are also transforming the squared error. Is this a problem? Or is it actually helping the model to find a better minimum error in the un-transformed space?

1 Like
28

so what does MAPE with neural network mean? I would have expected the result to be the same, but why is it lower (0.093)?

29

So the prefix is not fed into an RNN? Its just fed into a dense layer?

30

Is this a regression problem or classification? Are they predicting GPS coordinate numbers and error calculated as a distance from the destination?

1 Like
31

This p_i * c_i is very similar to what happens in the memory network paper (Babi memn2n). In that case, the output embeddings are weighted by the “attention” probability vector.

32

Can you talk more about the idea you have about first having the convolutional layer and passing that to an RNN? What do you mean exactly by that?

33

@thunderingtyphoons regression problem. yes, they’re predicting GPS coordinates

1 Like
34

So for the destinations we would have more error for the peripheral points? As we are taking a centroid of clusters?

35

Is Jeremy going to talk about segmentation tonight?

1 Like
36

For such types of problems(segmentation), what does the Y look like?
Is it the class number for each pixel for the whole image?

37

@layla.tadjpour the example he is currently doing is segmentation

38

Why crop it into 224x224? Are we using VGG for pre-training?

1 Like
39

@harveyslash yes, the output has a category number in the place of each pixel

1 Like
40

What about “The One Hundred Layers Tiramisu” paper ? Last week he said he would talk about that after Dense net.