/
KNOWN_BUGS
28 lines (17 loc) · 1.23 KB
/
KNOWN_BUGS
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
Code is only tested and known to compile on Debian/Linux (amd64) and
Windows 64bit platform. Minimal testing has been done on Intel 32bit plaform.
* t-SNE is a bit buggy(?) and don't scale now to large number of data points
* reinforcement code don't learn example cart-pole upswing properly
* VAE code (Variational Autoencoder) currently doesn't handle learning z-space model variance well.
is it possible to alter computation to use zvariance = encoder_stdev(x)*encoder_stdev(x) and handle related math correctly?
* HMC (bayes nnetwork) do not work properly on more that
one output dimension nnetwork? (recheck)
* OUT OF MEMORY problems with BIG neural networks or BIG DATA
leads to strange error messages that are not properly handled.
* ICA code probably has some bugs (with bad/singular data)
* static linking of nntool causes problems due to known gcc compiler bug
* realnumber (arbitrary precision math): sin() and cos() function calls fail
as they are not properly implemented.
* Sylvester Equation solver hasn't been working for a long time but is not needed (won't fix)
* matrix<T>::symmetric_pseudoinverse() don't handle complex numbers correctly.
symmetric_eig() to properly handle complex numbers so the code can be enabled.