Question: What is the optimal S parameter? about tmu HOT 3 CLOSED

Hi @BooBSD! The best s-value depends on the task, and different values yields different effects. A larger s provides more fine-grained patterns, while a smaller s may in some cases give more robust patterns (https://www.ijcai.org/proceedings/2022/616). I usually start with trying out 2.5, 5.0, 10.0 and 20.0, to learn more about what works well for the data. If you print out the clauses, you can also look at how they grow with increasing s. Did that clarify the role of s?

from tmu.

@olegranmo Thank you.
I tried to get the best result on MNIST classification and found out that I got better accuracy with S > 10 (up to 20) if I have more than 2000 clauses. And If I have 2000 clauses or less, for example 500, I got better accuracy if I use S ~= 10.

Is there some formula relationship between number of clauses and S? Like best T param for MNIST is sqrt(clauses_num/2).

from tmu.

Interesting! Finding such relationships, empirically and formally, are important steps ahead in the TM research. Feel free to report any findings here :-)

from tmu.