Timeline for My neural network 'learns' to estimate each class equally
Current License: CC BY-SA 3.0
11 events
| when toggle format | what | by | license | comment | |
|---|---|---|---|---|---|
| Apr 4, 2017 at 16:16 | answer | added | D.W. | timeline score: 1 | |
| Apr 4, 2017 at 13:32 | comment | added | M-- | @FraserOfSmeg I would try to debug that but meanwhile, if I were you, I would run an R script within vb.net. Look at here. And you cannot believe how easy it is to have a neural network in R. | |
| Apr 4, 2017 at 5:02 | comment | added | FraserOfSmeg | @Masoud thanks, I've added my code. If there's anything that I need to explain more please let me know (I'm bad at making pseudocode!). If my actual code would be helpful (it's written in vb.net) let me know! :) | |
| Apr 4, 2017 at 5:00 | history | edited | FraserOfSmeg | CC BY-SA 3.0 |
Added Pseudocode
|
| Apr 4, 2017 at 3:40 | comment | added | M-- | OK. I got what you mean. You need to let us see how you set up your model/code. @FraserOfSmeg | |
| Apr 4, 2017 at 3:33 | comment | added | FraserOfSmeg | @Masoud I'm not sure I understand what you're getting at. There are 3 classes of Iris, but for any single line of input my model predicts 0.33 for each class. It's not learning to differentiate between the classes. | |
| Apr 3, 2017 at 21:30 | comment | added | M-- |
Would you take a look at iris data-set. It is divided into 3 blocks by species column. Do you think that would help you to figure out why it should go to 1/3?
|
|
| Apr 3, 2017 at 14:57 | comment | added | FraserOfSmeg | The output of the network is a vector which should be near 0 everywhere but the index of the correct type of flower. (eg. 1,0,0 or 0,1,0 or 0,0,1). The network does start with random values (and as it happens initially predicts higher values (~0.98) for each class) then over time iterates down to 0.33... | |
| Apr 3, 2017 at 14:33 | comment | added | liangjy | I would check if your backpropagation implementation works, since outputting each class 1/3 of the time is what I'd expect if you randomly initialized the weights and never updated them. | |
| Apr 3, 2017 at 14:18 | comment | added | SmallChess | Are the actual predictions accurate? How do you pick the class for classification? | |
| Apr 3, 2017 at 13:29 | history | asked | FraserOfSmeg | CC BY-SA 3.0 |