Timeline for Feedforward neural network for sinusoidal prediction
Current License: CC BY-SA 3.0
12 events
| when toggle format | what | by | license | comment | |
|---|---|---|---|---|---|
| S Mar 28, 2018 at 11:20 | history | edited | Peter Flom | CC BY-SA 3.0 |
Fixed LaTeX formatting
|
| S Mar 28, 2018 at 11:20 | history | suggested | jojeck | CC BY-SA 3.0 |
Fixed LaTeX formatting
|
| Mar 28, 2018 at 10:49 | review | Suggested edits | |||
| S Mar 28, 2018 at 11:20 | |||||
| Mar 26, 2018 at 14:20 | comment | added | Peter Barrett Bryan | Relevant: pdfs.semanticscholar.org/05ce/… | |
| Mar 25, 2018 at 21:26 | comment | added | Peter Barrett Bryan | Excellent clarification of my misunderstanding. Thank you! | |
| Mar 25, 2018 at 21:23 | comment | added | Jakub Bartczuk | Ok I misunderstood that part. I've edited answer accordingly. | |
| Mar 25, 2018 at 21:22 | history | edited | Jakub Bartczuk | CC BY-SA 3.0 |
added 533 characters in body
|
| Mar 25, 2018 at 21:17 | vote | accept | Peter Barrett Bryan | ||
| Mar 25, 2018 at 21:16 | comment | added | Peter Barrett Bryan | Sorry, first link relates to RNNs/LSTMs use on periodic data (even if not strictly temporally serial). The second one is just an interesting extension. By establishing an effective identity between GPs and NNs of infinite width, a single layer NN of sufficient width should be able to handle the distribution. Thanks for an excellent answer. Got my neurons firing | |
| Mar 25, 2018 at 21:09 | comment | added | Jakub Bartczuk | Yes, the number of nodes is unbounded in general (I mean in general, because sometimes it might be possible to reconstruct function perfectly , for example when $f$ is just affine transformation composed with activation function). Interesting paper BTW, but I don't exactly get how it relates to RNNs | |
| Mar 25, 2018 at 21:04 | comment | added | Peter Barrett Bryan | The motivation for my mention of RNNs/LSTMs was their use on periodic data à la goelhardik.github.io/2016/05/25/lstm-sine-wave. It looks like folks have accomplished the prediction as in stackoverflow.com/questions/13897316/…. Interesting that the theorem extends to even one layer networks. Is the premise, then, that the number of nodes in the single layer is unbounded? I think in the limit there is proof that NNs of a single layer are upper bounded by gaussian process performance arxiv.org/abs/1711.00165 | |
| Mar 25, 2018 at 20:54 | history | answered | Jakub Bartczuk | CC BY-SA 3.0 |