From listmaster animalgenome.org Thu Aug 1 13:46:17 2019
Return-Path: <listmasteranimalgenome.org>
From: Ignacy Misztal <ignacyuga.edu>
Subject: RE: Guide on deep learning
Postmaster: submission approved by list moderator
To: Members of AnGenMap <angenmapanimalgenome.org>
Date: Thu, 01 Aug 2019 13:46:17 -0500
Very funny but relevant message by Daniel. Is SHALLOW DEEPly out of fashion?
Miguel Perez-Enciso visited UGA a couple of weeks ago and gave a short
course on machine learning. We had many discussions, and although Miguel is
capable of sophisticated theories, he adapts to new (good or bad) times.
Some 25+ years ago the animal breeding was full of sophisticated theories
with long derivations. Write a model, derive derivatives, estimate
parameters... Took a guru to derive anything, but there were many gurus at
that time. E.g., see formulas for threshold models (Gianola, Fouley,
Hoeschele,..) or for AI REML (Thompson, Jensen,..). Brrrrrr
Next came models based on sampling (sometimes termed Bayesian although BLUP
is Bayesian too) strongly advocated by Daniel Gianola. Write a model but
sample instead of deriving much. What needed pages of formulas could then
be described in few formulas for sampling. Combinations of
threshold/linear/censored models are now trivial to implement! Except that
costs are high, results need interpretation, and the methodology is not
really suitable for large evaluations. But good for papers and requiring
less class time for students.
The next step is machine learning. Do not even formulate a model, let
machine do it from a list of cryptic choices. Would you trust machine
learning in commercial decisions in animal breeding? Not sure but DEEP is
good for papers and requires even less class time. Also, DEEP is good when
you need to have results (of whatever quality) in seconds.
Recently, the US public television (PBS) produced a program "Nova:
Predictions by the numbers" that cited Fisher as originator of statistical
significance, also commenting on the Bayes theorem. The program was ranked
by viewers as 2 out of 5. Complexity seems to be out of fashion nowadays.
Ignacy Misztal
> -----Original Message-----
> From: DANIEL GIANOLA <gianolaansci.wisc.edu>
> Sent: Monday, July 22, 2019 12:11 PM
> To: Members of AnGenMap <angenmapanimalgenome.org>
> Subject: Re: Guide on deep learning
>
> A DEEP THANK YOU, Miguel and Laura. Hopefully, we will all learn DEEPly, and
> see if knowledge of the DEEP architecture of complex traits can ever be learned.
>
> We certainly cannot predict with the SHALLOW y=Xb+e (GWAS), and it is argued
> that if the whole-genome is implicated, then GWAS is not informative (Tam et al.
> 2019, Nat. Rev. Genetics). That line of reasoning implies that perhaps we cannot
> learn much with the less shallow SHALLOW y=Xb+Zu+e--all loci get
> implicated-- although predict we certainly can!
>
> Maybe DEEP learning will implicate the whole-genome, the whole interactome,
> the increasing army of DEEP phenotypes (images, sounds, odor--good and bad
> ones, whatsapp, etc), and eventually, a DEEP MULTI-OMICS SINGLE STAGE BLUP
> will become a focal topic in the next five years. Perhaps, Ignacy Misztal will
> extend APY to DEEPY and Cheng-Fernando-Garrick will develop DEEP Bayes Cpi-
> pipi-pipipi, etc. Here pi means first layer, pipi means the second layer, and
> pipipipi stands for higher-dimensional stuff.
>
> Another possibility is that we will drown in a DEEP sea of confusion, akin to the
> situation of choosing a beer in a good Belgian brasserie or extracting a
> meaningful message from the UN General Assembly.
>
>
> Regards,
>
> Daniel
>
> ________________________________
> .From: miguel.perez <miguel.perezuab.es>
> .Sent: Monday, July 22, 2019 8:02 AM
> .To: Members of AnGenMap <angenmapanimalgenome.org>
> .Subject: Guide on deep learning
>
> Just in case you are wondering about how deep learning works, we have posted
> an easy to follow script in github guide
> https://github.com/...rezenciso/DLpipeline
> and associated, somewhat more technical reference in Genes, 10, 553.
>
> https://www.mdpi.com/2073-4425/10/7/553
>
> --
> ===============================================================
> Miguel Perez-Enciso
> ICREA professor
> Centre for Research in Agricultural Genomics (CRAG) and Facultat de Veterinaria
> UAB Campus Universitat Autonoma Barcelona Bellaterra
> E-08193 Spain
> Tel: +34 935636600 ext 3346
> Fax: +34 935636601
> miguel.perezuab.es
> http://www.icrea.cat/...Enciso-255
> http://bioinformatics.cragenomica.es/numgenomics/
> http://scholar.google.es/...r=Lpl_-dcAAAAJ&hl=es
> http://orcid.org/0000-0003-3524-995X
> https://github.com/miguelperezenciso/
> ================================================================
|
