4/20
-
classical Horseshoe - will lead to very small weight, then near-zero kernel
-
solution: only use local scaler + softmax + block-diag (ensire K is diag)
-
works well on low freq ( 50 scale + sin(pi x) ), but not on x2 + sinx
- always result in all average weights (1/Q)
-
to be test:
- full cov +only local scaler + softmax?
-
-
whiting trick (not use log-det term in loss)
- help convergence on low freq ( 50 scale + sin(pi x) ),
- not help (even hard) other case
-
real whiting trick + VI