This newsletter made possible by MosaicML. Btw, we’re looking for early customers who spend a lot on training neural nets for computer vision or NLP in PyTorch on cloud GPUs.
Figure 1 of "Human-level Atari 200x faster" is not visually clear about the speedup of MEME. IMHO, authors shoud change the chart to be a "side-by-side bar chart" instead of "stacked bar chart". The way it is presented may give the interpretation that the total bar is a "sum" of the orange bar with the blue bar, which apparently it is not the case: my interpretation (for the legend works) is that the orange bar has the size of the blue bar plus an "additional orange size". That's really not trivial for a stacked bar chart.
Figure 1 of "Human-level Atari 200x faster" is not visually clear about the speedup of MEME. IMHO, authors shoud change the chart to be a "side-by-side bar chart" instead of "stacked bar chart". The way it is presented may give the interpretation that the total bar is a "sum" of the orange bar with the blue bar, which apparently it is not the case: my interpretation (for the legend works) is that the orange bar has the size of the blue bar plus an "additional orange size". That's really not trivial for a stacked bar chart.