Hacker News
new
|
past
|
comments
|
ask
|
show
|
jobs
|
submit
login
gugagore
on Sept 29, 2017
|
parent
|
context
|
favorite
| on:
Nonlinear Computation in Deep Linear Networks
I didn't understand how the gradients are produced to honor this underflow behavior. Is that the reason why they use "ES" instead of symbolic (or probably they meant automatic) differentiation?
dnautics
on Sept 30, 2017
[–]
correct.
Guidelines
|
FAQ
|
Lists
|
API
|
Security
|
Legal
|
Apply to YC
|
Contact
Search: