cool paper, i enjoyed following the post (and really appreciated the effort put into the bokeh viz!)
i wonder if this would have been improved by being clearer about the motivation. the authors frame it as though the ~60ms penalty for using the LSTM for prediction is a huge burden, and i can imagine situations where it is. however, it seems like if this is the case, we need some real life/"scaled out" examples of how this solution would work in practice. e.g. how long does the decision logic take to execute (maybe 5ms?); what proportion of the time will you have to run the LSTM after the BoW model anyway? note that those instances you are now worse off than just running the LSTM in the first place (total time = BoW time + decision time + LSTM time). once you have all these you can run the math and know (on average) how much time you'll actually save, and how much performance you sacrifice
In the Baseline section, you will find a sentence with the following: "However, there is a cost to using the strategies. We have to run all of the sentences through the bag-of-words model first, to determine if we should use the bag-of-words or the LSTM." followed by some math that takes the added cost into account, which is also what we based our plots/results on.
A 'batch' is how much of the data you put in memory at once while training the NN. To train even a small language model, you'll go through 1000s of batches, so the time difference is way bigger than it sounds. I agree a more practical example would have been nice -- maybe it'll come out in the paper.
It was based on test time prediction, so given you have received a sentence, how fast does it take to compute the prediction with either a bag-of-words or an LSTM.
When you say practical example, would that be in the scenario that you have an API server running? So to consider such costs as latency, data transfer, API overhead etc.?
i wonder if this would have been improved by being clearer about the motivation. the authors frame it as though the ~60ms penalty for using the LSTM for prediction is a huge burden, and i can imagine situations where it is. however, it seems like if this is the case, we need some real life/"scaled out" examples of how this solution would work in practice. e.g. how long does the decision logic take to execute (maybe 5ms?); what proportion of the time will you have to run the LSTM after the BoW model anyway? note that those instances you are now worse off than just running the LSTM in the first place (total time = BoW time + decision time + LSTM time). once you have all these you can run the math and know (on average) how much time you'll actually save, and how much performance you sacrifice