 |
.. (לתיקייה המכילה) | |
Can we assume that softmax will be used on the last layer only? | |
| Yes. |
The softmax is appeared to be numerically unstable, what can we do? | |
|
Subtract the max value from all the inputs to avoid overflow, see the link below for more info: https://stackoverflow.com/questions/42599498/numercially-stable-softmax |
How we should deal with multiple output regression? | |
|
You can assume that regression output is in R^1 as seen in class. |
Do we need to compute train loss and error at the end of every epoch? | |
|
No, it is sufficient to average the error and loss along training. (Yes it doesn't give the same result as evaluating at the end but it gives sufficient information and it is much more efficient). |
How to build the test-set for the regression? Do we need to do Cartesian product? | |
|
Yes, You should sample a grid in 2D with equal spaces. If 1000*1000 is to large you can do 200*200 points. |