I'm new in DeepLearning. I'm not good at understanding and commenting on graphics.Can you help me with these graphs 
$\begingroup$ $\endgroup$
2 
- $\begingroup$ Is this question homework? $\endgroup$noe– noe2021-05-01 21:27:55 +00:00Commented May 1, 2021 at 21:27
- $\begingroup$ No I just worked making something in a dataset $\endgroup$Sdeveci– Sdeveci2021-05-01 21:29:26 +00:00Commented May 1, 2021 at 21:29
Add a comment |
1 Answer
$\begingroup$ $\endgroup$
3 The questions I ask myself when I see learning graphs are the following ones:
- Is the loss decreasing and the accuracy increasing ? if yes, your network is learning and everything works fine, which is already a good new.
- Have we reached a kind of plateau ? (in accuracy especially), which means learning is over. (Here maybe you could train your network a bit more, it doesn't seem to have reached a plateau yet).
- What is the value of that plateau ? Final validation accuracy is the most important value we seek when learning.
- Do the curve have an exponential shape ? This is a tricky one and i'm not sure everyone does this, but i like having exponential shape curves when learning (which is not possible with every optimizer), so in your case (linear curves) i would try increasing the learning rate to get a faster learning.
- Is the validation accuracy decreasing at the end of the training ? in which case it means we are overfitting and need to stop learning earlier. (Not your case here)
Hope this helps, this is only my tips for analysing learning graphs and may not be what everyone ask theirself when facing these graphs.
- $\begingroup$ Doesn't it seem like it is hitting a plateau at around the 50 epoch mark? $\endgroup$Jayaram Iyer– Jayaram Iyer2021-05-02 03:37:58 +00:00Commented May 2, 2021 at 3:37
- $\begingroup$ @JayaramIyer Well it definitly slows down, but we still improve accuracy for about 5% on the next 50 epochs so accuracy is still improving, the slow improvement is more likely to be due to low learning rate rather than learning being finished imo. $\endgroup$Ubikuity– Ubikuity2021-05-02 12:20:45 +00:00Commented May 2, 2021 at 12:20
- $\begingroup$ Perhaps try an adaptive optimization algo such as Adam. $\endgroup$Jayaram Iyer– Jayaram Iyer2021-05-02 12:59:37 +00:00Commented May 2, 2021 at 12:59