It is good for both.
2 Likes
It is also good for regression.We will see multiple examples of that.
1 Like
What is your take on Deep Learning models for Information Retrieval ?
Dear Amazon, I bought a toilet seat because I needed one. Necessity, not desire. I do not collect them. I am not a toilet seat addict. No matter how temptingly you email me, I’m not going to think, oh go on then, just one more toilet seat, I’ll treat myself.
17 Likes
A really depressing example of this is someone that purchased an urn after the death of a loved one, and then was recommended more urns for months afterwards.
13 Likes
It’s a good thing to do because when you have a head-only part, which usually is bunch of dense layers, you want the weights to converge to a decent range where the predictions are as good as possible. The you allow all of the network to train but at a much slower learning rate. If you look at the source code for fine_tune, you do one-cycle training on head followed by unfreeze and another one-cycle training but at 1/100 learning rate. The later steps stabilizes and changes the weights based on the data.
1 Like
The problem with this data set is that it is so relatively sparse at lower temperatures. I would worry that the relatively small number of points at the lower temperatures have too much leverage on the slope. I would not be inclined to believe that the result is different than the null hypothesis.
1 Like
FYI: daily effective reproductive
number = R
2 Likes
I’m assuming that paper has been peer reviewed. Was this something that was caught, or is this paper being accepted?
1 Like
What are the numbers in the bracket? (on the slide below the equation)?
It seems like this aside might be more appropriate for the COVID video that comes out of tonight’s lecture, instead of the ML portion.
1 Like
Jeremy has been defending this paper on Twitter so unless he has changed his opinion on this paper since then I’m expecting a plot twist at the end of this explanation 
EDIT: to be clear, you probably still can’t draw proper conclusions on the graph alone, but the graph isn’t the whole paper
6 Likes
I’ve heard plenty of problems with p-hacking. Making p fit your assumptions.
3 Likes
I wouldn’t classify this as an aside, as these topics are important for understanding a lot of scientific research.
5 Likes
Those numbers are the uncertainties (standard deviations) on the fitted intercept and slope
2 Likes
The approach of teaching seems very different this time compared to last time. There is less talk of DL and more general conversation in modelling. Is this intentional?
6 Likes
Isn’t there another statistic, beta?, that measures the discriminative power of the test? Would that with p help determine whether the data show an effect or not?
1 Like
p-values guide quite a lot of research in biology and medicine still. Unfortunately so.
Lots of research dollars are spent to produce something significance of which is determined by p-value!
No equivalent, easy & popular alternatives exist though.
2 Likes