Timeline for Why is my neural network not working?
Current License: CC BY-SA 3.0
19 events
when toggle format | what | by | license | comment | |
---|---|---|---|---|---|
Nov 30, 2016 at 15:20 | history | edited | viceriel | CC BY-SA 3.0 |
added 239 characters in body
|
Nov 29, 2016 at 16:12 | history | undeleted | Matt | ||
Nov 29, 2016 at 15:41 | history | edited | user1228 | CC BY-SA 3.0 |
You most likely had this answer deleted because you implied it should have been a comment. But this is definitely worth an answer. I'm editing to fix layout and remove that comment (fyi, two spaces at the end of a line == newline). Also flagging to undelete
|
Nov 28, 2016 at 17:44 | history | deleted | Martijn Pieters | via Vote | |
Nov 28, 2016 at 14:32 | review | Low quality answers | |||
Nov 28, 2016 at 15:22 | |||||
Nov 28, 2016 at 13:48 | comment | added | viceriel | I don't know. But if you can test your BP algorithm, use your BP on XOR problem. Net parameters can be 2 2 1 | |
Nov 28, 2016 at 13:36 | comment | added | harry lakins | it takes 30 secs to train one epoch with that many! How come in the tutorial i linked, they get a 98% success within just minutes from using 784,15,10? | |
Nov 28, 2016 at 13:32 | comment | added | viceriel | Ok, once again. Net trained by better approach using on this classification totally 4 000 hidden units, your net using 80. Yes, more neurons more slow :/ | |
Nov 28, 2016 at 13:26 | comment | added | harry lakins | having two hidden layers of 40 neurons each results in an error of 2.24999 and it never improves. It is also dramatically slower (have tried with many different learning rates). Im pretty sure there is something wrong with my actual aglorithm | |
Nov 28, 2016 at 12:14 | history | edited | viceriel | CC BY-SA 3.0 |
added 59 characters in body
|
Nov 28, 2016 at 12:13 | comment | added | harry lakins | Will do - and will get back to you - how many neurons on each so you suggest? | |
Nov 28, 2016 at 12:12 | comment | added | viceriel | Yes, use two hidden layers and much more hidden neurons. | |
Nov 28, 2016 at 12:12 | history | edited | viceriel | CC BY-SA 3.0 |
added 254 characters in body
|
Nov 28, 2016 at 12:10 | comment | added | harry lakins | Are you saying you do not recommend that structure ? | |
Nov 28, 2016 at 12:09 | comment | added | viceriel | Deep net resolving same task has parameters: 1. 768 2. 500 - first hidden | |
Nov 28, 2016 at 12:09 | comment | added | viceriel | And maybe. You try classificate a handwriting characters. 768 input units, 15 hidden, 10 output. | |
Nov 28, 2016 at 11:47 | comment | added | viceriel | In short, Gradient vanishing problem. You compute gradient on output units and then on hidden layers units and than on hidden layer units and than.... and than you have zero error signal and first layers aren't learning anything. | |
Nov 28, 2016 at 11:38 | comment | added | harry lakins | I have already tried -1,1. Which have even worse results. What do you mean about the hidden layer count and the ability of back prop? Edit answer if you can't comment :) | |
Nov 28, 2016 at 11:35 | history | answered | viceriel | CC BY-SA 3.0 |