# Showing standard deviation for training curve

I am training a neural network and I wanted to plot the evolution of different metrics (MSE…) during training. To get an idea of the variations between between different trainings, I am using several models and plotting average value and standard deviation. My problem is the following: I did not manage to find a good way to plot this curve and also did not find any good explanation on internet.

Let's denote by y the metric to be displayed, and x is the number of epochs. On the plots below, one standard deviation is given by the shaded area.

Here is what I tried and why I think it is not working:

1. A linear scale is not great because the metric spans different order of magnitude. In particular, the early values are much larger than late values. One possibility could be to start the curve at epochs 5, for example, to remove the early large values, but I would prefer to display everything.

2. A log scale seems much more natural in view of the previous point, however variations can be large, of similar scale compared to the scale of the mean value. Since the space between two ticks is geometrically distributed, one standard deviation below the curve appears much bigger than above. This tends to produce large spikes on the figure, or even completely fill the area below the curve.

1. I have tried to display the relative standard deviation instead (rescaled by $$\log_{10} \mathrm e$$). This is the appropriate standard deviation for the variable $$z = \ln y$$. However, I dont want to plot z but really y in log scale. Nonetheless, I thought it's fine to plot the relative standard deviation if it displays better. But, I am not sure if it's really intuitive what it shows. A worse problem is that the standard deviation appeard very large for small values (since we divide by a number < 0).

1. Finally, I tried to use the symlog scale but I am not sure it's a good solution: if the linear threshold is large, then I find that that it makes the figure difficult to interpret. If it is small, we get the same problems as for the log scale. Below is a plot for a threshold of 0.01.

Conclusion: what would be the best way to display the standard deviation a log scale?

• How about linear scale but show two- or three-times standard deviation? Maybe not standard but as long as it's clearly labelled in the plots, it might show enough to allow people to see the relative spread of the data across the epochs. Dec 25, 2020 at 21:58
• Thanks for the suggestion. This would indeed help to see the standard deviation better on the firstp lot, but the mean value is still not readable. For example with learning rate decay (and longer training), you would see a succession of plateau in log scale, but not in linear scale. Dec 26, 2020 at 1:16
• Perhaps two plots then? One with a multiple of standard deviation with a linear scale and then another with log to show the learning rate? Otherwise, I think I'd go with the log scale – sure it has some drawbacks, as you've described, but it's probably the most common plot type after linear and I think more intuitive than options 3 and 4. Dec 26, 2020 at 11:26