TastyPancakes
/
jwebsite
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

[ADD] New adventure added

This commit is contained in:
Eduardo Cueto-Mendoza 2021-04-08 20:29:56 +01:00
parent 7898128359
commit 7cfa202520
1 changed files with 19 additions and 2 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

21
index.md
View File

@ -28,9 +28,26 @@ function len(p::T) where T<:Point
end
```
## 07/04/2021
OK it is been a while since I wrote on the blog
## Arriving to Dublin
OK it is been a while since I wrote on the blog, due to the (COVID-19) pandemic, the projects were maintained by me are on stand by and in
need for an umdate now that Julia is on version 1.6.0 (it was 1.4.3 when they started) but I'm still working hard now into a more broader
than just compressing a NN architecture. Now, models are so big their number of parameters is huge, just look at GPT-3; this has opened new
questions on the area specifically how "efficient" are the compression techniques when applied to those huge models.
Thanks to those theoretical advances now the previous problem of reducing the size of a model has become a multivariable optimization
problem, where other hyperparameters influence the "efficiency" of how a model performs both during training, and inference. As an example,
we can mention the amount of FLOPS performed, the amount of energy consumed in Watts, clock time, etc. (all during training and inference).
This means that currently I am surveying SOTA models to compare if their gains on accuracy/precision are indeed valuable with respect the
previous, for example if the precision of a new model increments 0.0001 than former models but it consumes 100 times the amount of energy or
then that architecture is not efficient and should only be used on applications where that gain might be required. Also, I am exploring
ideas from Topological Data Analysis to try and find smaller architectures that have as good performance and generalization as their more
parametrized counterparts.
Regarding my move to Dublin was a hassle...to arrive I had to take 3 flights one to Mexico City, then one to Dallas and finally to Dublin
at least it wasn't as expensive due to COVID. I haven't been able to travel in Ireland, still on lockdown (08/04/2021) and might continue for
at least a month.
## A bit about myself