Timeline for What should I do when my neural network doesn't learn?
Current License: CC BY-SA 4.0
16 events
| when toggle format | what | by | license | comment | |
|---|---|---|---|---|---|
| Jun 24, 2021 at 3:24 | comment | added | WestCoastProjects | This is actually a more readily actionable list for day to day training than the accepted answer - which tends towards steps that would be needed when doing more serious attention to a more complicated network. | |
| Jun 1, 2021 at 8:47 | comment | added | Imran Kocabiyik | Point 1 is also mentioned in Andrew Ng's Coursera Course: coursera.org/learn/… I previously had the issue you mentioned in point 6. It is called "Training-Serving Skew": developers.google.com/machine-learning/guides/… | |
| May 13, 2021 at 21:50 | comment | added | A Tyshka | @Alex R. I'm still unsure what to do if you do pass the overfitting test. In my case it's not a problem with the architecture (I'm implementing a Resnet from another paper). Although it can easily overfit to a single image, it can't fit to a large dataset, despite good normalization and shuffling. Likely a problem with the data? | |
| Aug 5, 2020 at 18:10 | comment | added | Azmisov | Testing on a single data point is a really great idea. If it can't learn a single point, then your network structure probably can't represent the input -> output function and needs to be redesigned. | |
| S Oct 14, 2018 at 16:39 | history | suggested | Matti Wens | CC BY-SA 4.0 | Typos, consistent formatting, improvements to English usage. |
| Oct 14, 2018 at 16:19 | review | Suggested edits | |||
| S Oct 14, 2018 at 16:39 | |||||
| Sep 29, 2018 at 19:11 | history | edited | Alex R. | CC BY-SA 4.0 | deleted 2 characters in body |
| Jun 20, 2018 at 23:49 | history | edited | Alex R. | CC BY-SA 4.0 | deleted 10 characters in body |
| Jun 20, 2018 at 13:41 | comment | added | John Coleman | Making sure that your model can overfit is an excellent idea. I am so used to thinking about overfitting as a weakness that I never explicitly thought (until you mentioned it) that the ability to overfit is actually a strength. | |
| S Jun 20, 2018 at 5:24 | history | suggested | CommunityBot | CC BY-SA 4.0 | Fixed numbering |
| Jun 20, 2018 at 4:32 | review | Suggested edits | |||
| S Jun 20, 2018 at 5:24 | |||||
| Jun 19, 2018 at 19:00 | history | edited | Alex R. | CC BY-SA 4.0 | added 6 characters in body |
| Jun 19, 2018 at 18:54 | history | edited | Alex R. | CC BY-SA 4.0 | added 195 characters in body |
| Jun 19, 2018 at 18:47 | history | edited | Alex R. | CC BY-SA 4.0 | added 195 characters in body |
| Jun 19, 2018 at 18:47 | comment | added | Sycorax♦ | (+1) Checking the initial loss is a great suggestion. I regret that I left it out of my answer. | |
| Jun 19, 2018 at 18:45 | history | answered | Alex R. | CC BY-SA 4.0 |