Lesson 4 In-Class Discussion

Part 1 (2018)
81

Any suggestions on Linear - Relu - BN vs Linear - BN - Relu (like the BN paper suggests)? Is it because one works better than other?

1 Like
82

Yeah I agree, basically you have to assume what you can change and about how much you can change, for the labels to still be the same.

83

My guess is school holiday and promo are already hot coded because it is Boolean.
other categorical variables are going to become hot coded.

3 Likes
84

That log rule is not right. log(a/b) = log(a) - log(b) which not the equal to log(a)/log(b)

3 Likes
85

Are you installing feather-format? This works for me on Crestle:

pip3 install feather-format
2 Likes
86

Wow thanks for the insight!

87

True was just typing that great catch :slight_smile:

88

I owe you a beer :slight_smile:

1 Like
89

It’s not really the same. For normalization you use only the data but for binning you choose arbitrary values. Most of the time these are driven by domain specific knowledge.

90

would dropout in this case put certain columns to 0 or would you take out the entire row?

91

certain columns

2 Likes
92

¯\(ツ)/¯ sounds right to me.

93

Why does 2014 get converted to 2? Is it because there’s only 2012 and 2012 is 0, 2013 is 1, and 2014 is 2?

94

Is he describing one-hot encoding currently?

95

embeddings.

1 Like
96

He is explaining embeddings.

1 Like
97

@yinterian just a note to ask jeremy to tell us the trick (at 53:00 minute) towards the end of the class. :sweat_smile:

1 Like
98

2012 is the starting year inside the dataset.
2013 is the next one and so on…

99

Sunday was a Rank1 matrix of length 4 (4x1). How could it fit (get appended) into the original input, which was nx1. I did not understand the dimensions.

3 Likes
100

So do you not use one-hot encoding when doing this at all? It just uses the embedding matrix instead?

1 Like