Edit ‘scaling_laws’
This commit is contained in:
parent
05f018be1f
commit
a3df8f47b9
@ -1 +1 @@
|
|||||||
The principle on which scaling laws are based is that straight lines on graphs are the most powerful force in the universe. If a sufficiently smooth trendline (on a linear or log scale) is discovered, the universe will contrive to extend this trend line (until it doesn't; see [[Moore's Law]]). Scaling laws are important to [[biology]] (as organisms are constrained by limits such as heat dissipation, gas/chemical exchange across membranes or via lungs, mechanical support, and power consumption and will expand up to those limits) and [[deep learning]], in which metrics such as [[training loss]] usually scale smoothly according to power laws, due to the [[fundamental apiostructure of reality]].
|
The principle on which scaling laws are based is that straight lines on graphs are the most powerful force in the universe. If a sufficiently smooth trendline (on a linear or log scale) is discovered, the universe will contrive to extend this trend line (until it doesn't; see [[Moore's Law]]). Scaling laws are important to [[biology]] (as organisms are constrained by limits such as heat dissipation, gas/chemical exchange across membranes or via lungs, mechanical support and power consumption and will expand up to those limits) and [[deep learning]], in which metrics such as [[training loss]] usually scale smoothly according to power laws, due to the [[fundamental apiostructure of reality]].
|
Loading…
Reference in New Issue
Block a user